# MrJev

> Independent guides, pricing, and a curated project directory for Jev, TypeSafe AI's System One model for typed decisions. MrJev is an independent, unofficial resource. Not affiliated with or endorsed by TypeSafe AI.

Jev is TypeSafe AI's System One model: it reads a state and answers typed questions — Choice (pick one option), Score (rate on ordered levels) and Noul (yes/no) — with a probability for every option instead of generated text. List price Jev 1.13: $0.042 per million input tokens, output not billed (checked 2026-09-18; see https://mrjev.com/pricing/ for current limits).

MrJev has 273 hands-on project pages (245 full reviews, 28 notes) out of 531 projects in the list, and 25 changelog entries. Generated 2026-10-12.

How we work:
- Reviews are hands-on: we install the project at a stated commit, run its tests, and drive it against local stand-ins that log every request, so we can say what it sends and how it fails. Each page names the commit and date tested.
- "Note" pages are short checks; "review" pages are full write-ups.
- We do not publish our own measurements of Jev's accuracy, latency or cost, and we do not repeat a project's own Jev figures. Measurements of open models are published with the hardware and method used.
- When a maintainer fixes something we reported, we retest and update the page; the issue we filed is linked from it.

To cite MrJev, link the specific page; review pages state the commit and date they describe. The full text of every page below is in https://mrjev.com/llms-full.txt.

## Start here

- [What Is Jev AI? The System One Model Explained](https://mrjev.com/what-is-jev/): Jev is TypeSafe AI's model for typed decisions: choices, scores, and calibrated probabilities instead of text. How it works, what it costs, and its limits. (updated 2026-09-18)
- [How to Use Jev: Getting Started with the Jev API](https://mrjev.com/guides/getting-started/): Jev tutorial: get early access, install the Python or JavaScript SDK, make your first Choice, Score, and Noul call, and act on confidence. (updated 2026-09-18)
- [Jev Confidence Thresholds: How to Choose Them](https://mrjev.com/guides/confidence-thresholds/): How to pick confidence thresholds for Jev on your own data: build a labeled set, measure accuracy by confidence, set risk-based cutoffs, and watch for drift. (updated 2026-09-18)
- [Jev Pricing Explained: $0.042 per Million Tokens](https://mrjev.com/pricing/): How Jev pricing works: $0.042 per million input tokens, free output, rate limits, and a worked comparison with Claude and GPT for high-volume classification. (updated 2026-09-30)
- [Best Jev Tools, Reviewed and Tested (September 2026)](https://mrjev.com/best-jev-tools/): We reviewed 240 popular Jev projects and tested 231 hands-on: what each is best for, what it sends where, and what to check before you install it. (updated 2026-09-22)
- [Jev Alternatives: 6 Ways to Get Typed AI Decisions](https://mrjev.com/jev-alternatives/): Jev alternatives for classification and routing: LLM structured output, small LLMs, fine-tuned classifiers, embeddings, and open-source reproductions. (updated 2026-09-19)
- [Jev Projects Directory](https://mrjev.com/projects/): Every community project in the MrJev/awesome-jev list, by category, with stars and hands-on reviews where we have them.
- [Jev Changelog](https://mrjev.com/changelog/): Official Jev and TypeSafe AI changes — models, pricing and limits, SDKs, docs — each linked to its source.
- [Jev Cost Calculator](https://mrjev.com/tools/cost-calculator/): Estimate monthly Jev cost from calls per day and tokens per call, next to frontier LLM list prices.
- [Jev Playground](https://mrjev.com/playground/): Build a Jev request in the browser and see the request body, answer shape and confidence.

## Latest official changes

- [An enterprise page with SOC 2, zero data retention and higher limits on request](https://mrjev.com/changelog/#2026-10-10-enterprise-page): 2026-10-10. typesafe.ai/enterprise is new, along with a contact page.
- [TypeSafe raises an $870M Series A at a $7.5B valuation](https://mrjev.com/changelog/#2026-10-10-series-a): 2026-10-10. TypeSafe announced its Series A, titled "TypeSafe A raises Series AI": $870 million at a $7.5 billion valuation, led by Andreessen Horowitz with Sequoia Capital, existing investor DCVC, and angel investors.
- [OpenRouter now lists Jev with a 64K context](https://mrjev.com/changelog/#2026-10-09-openrouter-64k-context): 2026-10-09. OpenRouter changed the context length it lists for typesafe/jev-1.13 and ~typesafe/jev-latest from 32,000 to 64,000 tokens, with up to 57,600 completion tokens.
- [TypeSafe publishes a customer case study, from Jack & Jill](https://mrjev.com/changelog/#2026-10-08-jack-jill-case-study): 2026-10-08. The TypeSafe blog published a case study: Jack & Jill, an AI-agent talent marketplace, replaced Gemini 3.1 Flash Lite with Jev for the first-pass candidate scoring step in its matching pipeline, then used Jev in other parts of the product.
- [The files the cookbooks read are now public](https://mrjev.com/changelog/#2026-10-06-cookbook-files-published): 2026-10-06. typesafe-ai/typesafe-public-examples is a new repository in the official organisation, MIT-licensed, described as "content used as part of documentation, demos, and cookbooks".
- [TypeSafe publishes how confidence is calculated, including Score's](https://mrjev.com/changelog/#2026-10-03-confidence-formulas): 2026-10-03. The confidence page gained a section called How confidence is calculated, with an exact formula for each answer type:
- [Jev's rate limits are cut to 100K tokens/s, and the request limit moves to per-second](https://mrjev.com/changelog/#2026-09-30-rate-limits-change): 2026-09-30. The model page changed one table row for jev-1.13.0:
- [TypeSafe publishes the code behind evals.typesafe.ai](https://mrjev.com/changelog/#2026-09-30-workflowevals-published): 2026-09-30. typesafe-ai/WorkflowEvals is a new public repository in the official organisation: the workflow code behind evals.typesafe.ai.
- [Python SDK 0.7.2 adds HTTP/2, and the base-URL rule gets a spec](https://mrjev.com/changelog/#2026-09-26-python-sdk-0-7-2): 2026-09-26. typesafe-sdk 0.7.2 is on PyPI with one new thing: an http2 extra, pulling httpx2[http2], and a documented way to use it.
- [Jev Router on OpenRouter picks the model and the effort for each request](https://mrjev.com/changelog/#2026-09-25-jev-router-openrouter): 2026-09-25. OpenRouter lists typesafe/jev-router, "TypeSafe: Jev Router", created on 25 September.
- [The Acceptable Use Policy is published, and the agreement binds to it](https://mrjev.com/changelog/#2026-09-23-acceptable-use-policy): 2026-09-23. The pointer that came and went last week is back, and this time the page behind it exists.
- [The cookbooks get an index page, sorted by technique and level](https://mrjev.com/changelog/#2026-09-23-cookbooks-index): 2026-09-23. docs.typesafe.ai/cookbooks is a new page: the seventeen existing cookbooks, grouped into Self-consistency, Batching, How-to, Extraction and Classification, each row labelled Beginner, Intermediate or Advanced.

## Reviews: Open Models & Reproductions

- [Laya MPS Note: Laya on a Mac, Three Memory Sizes](https://mrjev.com/projects/afshinm-laya-mps/): Serves Laya's typed-decisions model on Apple Silicon in three memory presets. Checked on CPU, not fully reviewed: all presets give identical answers. (note, tested 2026-10-07 at cdbf19e)
- [Open Jev Review: Train Your Own 150M Engine](https://mrjev.com/projects/intikhab49-open-jev-typed-decision-engine/): A 150M encoder with marker tokens and grouped softmax, trainable on a free T4. Since 6 October a checkpoint ships, and we reproduced its numbers on CPU. (review, tested 2026-10-07 at 78d3b3a)
- [sys1 Note: Laya Behind a Rust Server](https://mrjev.com/projects/alvarobartt-sys1/): A Rust server for Laya with a batching queue and Jev's /v1/systemone. Checked, not fully reviewed: since 75342d5 both official SDKs work unchanged. (note, tested 2026-10-06 at 75342d5)
- [Valen Review: Jev's Answer Shape, With Vision](https://mrjev.com/projects/liuziyu77-valen/): An open 2B multimodal decision model with Jev's answer types and formulas. Its one public checkpoint is a Sokoban specialist, weak on general yes/no. (review, tested 2026-10-06 at 3b764c9)
- [NotJev Note: One Token, a Real Probability](https://mrjev.com/projects/9pings-notjev/): Turns any OpenAI-compatible model with logprobs into typed decisions. Checked, not fully reviewed: the maths holds, and since v0.2.3 the official SDK reads it. (note, tested 2026-10-05 at ec50d98)
- [snap Note: Ask Everything, Generate Nothing](https://mrjev.com/projects/emnlmn-snap/): A llama.cpp-based decision server whose /v1/systemone works with the official SDK unchanged. Checked, not fully reviewed: the wire, and the CPU cost. (note, tested 2026-10-04 at bb1e6a2)
- [jevos Review: Yes/No Decisions on a Laptop CPU](https://mrjev.com/projects/feder-cr-jev/): A 1B decision model behind Jev's wire format, now a C++ server with real batching and a state cache. Negated questions are better than before, not yet sound. (review, tested 2026-10-04 at d48b9b7)
- [Jeff Review: A 0.8B Model That Beats Its Fallback](https://mrjev.com/projects/firelex-jeff/): Task adapters on a 0.8B decision model, with unusually reproducible evidence. Its headline holds — but the 27B 'fallback' is almost never used. (review, tested 2026-10-04 at d0173b4)
- [ollaya Review: Ollama, for Decision Models](https://mrjev.com/projects/ollaya-dev-ollaya/): Pulls open decision models by name and serves them behind Jev's wire format. A Jev MCP client drove it unmodified, including batch mode. (review, tested 2026-09-28 at 32acb6d)
- [CLM Review: The Artifact Beats the Headline](https://mrjev.com/projects/contrastive-lm-clm/): An 8B contrastive decision model behind a TypeSafe-shaped API. Its published run recomputes from its records, and says more than its README does. (review, tested 2026-09-27 at bb42c6c)
- [jevify Review: Jev's Shape From a Model You Have](https://mrjev.com/projects/fidecastro-jevify/): Turns any model you already serve into typed answers in Jev's response shape, and refuses the questions a backend cannot honestly answer. (review, tested 2026-09-25 at 1c0a0f9)
- [JevForge Review: A Server That Keeps Everything](https://mrjev.com/projects/zwlijay-jev-forge/): Trains, calibrates and serves an 0.8B decision model behind the Jev wire format — and writes every request it answers to disk, by design. (review, tested 2026-09-25 at 6cf0dfe)
- [JevK5 Review: An Open Jev Shape, and Sixteen Slots](https://mrjev.com/projects/allebee-jevk5/): An open-weight decision model that serves the /v1/systemone shape. Its server entry point needs CUDA, and a seventeenth option drops the connection. (review, tested 2026-09-24 at 01d3cc9)
- [OpenJev-Vision Review: One Image, Many Questions](https://mrjev.com/projects/iambusy-openjev-vision/): Encodes an image once and answers several typed questions from the shared posterior, offline on CPU. Its documented install cannot run its own test suite. (review, tested 2026-09-24 at b83bec9)
- [AgentJev Review: Typed Decisions on a Laptop CPU](https://mrjev.com/projects/malevrigns-agent-jev/): An open 0.6B decision model with a strict request contract, a loopback-only server and a checkpoint hash in /health — and one unit test that cannot import. (review, tested 2026-09-24 at a965ca8)
- [Rizzo Flow Review: Receipts All the Way Down](https://mrjev.com/projects/rizzo-ai-academy-rizzo-flow/): A local decision server on Spark-X2.5-4B. Every published number recomputed from its own rows, and every result file names its dataset by hash. (review, tested 2026-09-24 at d34665b)
- [OpenJevPro Review: Re-run, and Now It Adds Up](https://mrjev.com/projects/zhangcy122-openjev/): A 599-line client that asks an LLM to write its own scores. After our report it was re-run symmetrically, and the new table recomputes from its own records. (review, tested 2026-09-24 at 6f8564d)
- [Open Spark Jev Review: It Published Its Leak](https://mrjev.com/projects/abhishek085-open-spark-jev/): A local decision model for DGX Spark whose evaluation protocol names the time its own corpus leaked 97-99% of test states into training. (review, tested 2026-09-23 at f3bde93)
- [decider Review: An Open Model That Never Writes](https://mrjev.com/projects/mapika-decider/): An Apache-2.0 reproduction of the System One idea: typed questions in, a probability per option out, and a limits section that lists what it cannot do. (review, tested 2026-09-23 at 104b844)
- [Laya Review: The Limits Section Is The Draw](https://mrjev.com/projects/nandhakishorm-laya/): An open multilingual decision model that answered a complaint correctly in seven languages — and tells you, in its own benchmarks, where it does not hold up. (review, tested 2026-09-23 at 573e5b6)
- [Von Review: A Fast Local Model, Now Installable](https://mrjev.com/projects/wfzyx-von/): The option-marker path scored 93.6% on a third-party suite in our hands. The published weights answered at random for a day; that is fixed. (review, tested 2026-09-23 at a94aa36)
- [DeepOpen Review: Laya, Renamed and Trimmed](https://mrjev.com/projects/deepopen-com-deepopen/): Presented as an open decision engine with its own checkpoints. It is Laya 0.3.5 with edits, and the edits removed the part that routes non-English text. (review, tested 2026-09-22 at 28f7221)
- [JevBERT Review: A Compatibility Claim, Withheld](https://mrjev.com/projects/hawkymisc-typed-decision-bert/): A Jev-shaped server whose own compatibility document says not to call it compatible yet. The official SDK drove it with only a base-URL change. (review, tested 2026-09-22 at f0994cd)
- [Jev Visual Review: The Numbers Recompute](https://mrjev.com/projects/hr98w-jev-visual/): Many typed questions about one image in a single pass on Apple Silicon — and every figure in its results file regenerates from the records it ships. (review, updated 2026-09-22)
- [solar-mini4-jev: Every Number Checks Out](https://mrjev.com/projects/hunkim-solar-mini4-jev/): A wrapper putting Upstage's Solar Mini4 behind Jev's API shape. Its benchmark reproduces exactly — and it will forward whatever key you send to Upstage. (review, tested 2026-09-22 at fa401b5)
- [open-alternative-jev: One Pass, Every Question](https://mrjev.com/projects/ikermoel-open-alternative-jev/): Turns any open-weights LLM into a typed decider by reading one forward pass. A library, not an API replacement, and now with CI. (review, tested 2026-09-22 at 6ad87d7)
- [reflex Review: A Decision Model With No Weights](https://mrjev.com/projects/kshetrajna12-reflex/): An open re-creation of the System One idea that trains nothing: a frozen open-weights model, scored over your option set, behind one HTTP endpoint. (review, tested 2026-09-22 at dec50eb)
- [JEV-CPU Review: One Line Was The GPU](https://mrjev.com/projects/leesk212-jev-cpu/): A CPU port of SemIf that swaps the loader and nothing else. It ran offline on a laptop CPU in about a second — with the accuracy a 0.6B model buys. (review, tested 2026-09-22 at b49b5bf)
- [Dev-0.4B Review: The Numbers Match, The Demo Doesn't](https://mrjev.com/projects/mpnikhil-dev-0-4b/): A 399M encoder answering typed questions in one forward pass. Every published figure traces to a shipped artifact; the flagship example does not. (review, tested 2026-09-22 at 2ce2563)
- [AnyJev Review: Reverse the Options, Keep the Answer](https://mrjev.com/projects/nokia-applied-research-anyjev/): Nokia and Tencent researchers fix what reading option logits gets wrong. We reversed the options: the raw readout moved 3 of 6, theirs moved none. (review, tested 2026-09-22 at 78ca550)
- [choosekit Review: Forced Decoding, Zero Deps](https://mrjev.com/projects/notxf1le-choosekit/): Reads option probabilities straight out of your own llama.cpp. The fastest thing here, and its benchmark can attribute results to the wrong model. (review, tested 2026-09-22 at 5f1099e)
- [Dohnuts Review: A Model Card That Loses](https://mrjev.com/projects/psiace-dohnuts/): Small multimodal models for direct decisions, whose card publishes the benchmark where Jev beats it and lists what its own numbers cannot support. (review, tested 2026-09-22 at e22b01c)
- [OpenThai-SystemOne Review: Thai, One Pass](https://mrjev.com/projects/iapp-technology-openthai-systemone/): A 0.8B Thai and English decision model with a 256-slot head. It ran on four CPU cores, and the official SDK talks to its server unchanged. (review, tested 2026-09-21 at 849bce6)
- [PlayJev Review: Ten Games, One Forward Pass](https://mrjev.com/projects/omnijev-playjev/): A 0.8B model that plays browser games from pixels, with the harness that trained it. The games run anywhere; the model needs a GPU. (review, tested 2026-09-21 at 2d7a028)
- [Laya for Node: A Local Decision Model](https://mrjev.com/projects/receptron-laya/): Runs the open Laya model through ONNX from TypeScript, with the official request shape. Answers arrive in under a second on a plain CPU. (review, tested 2026-09-21 at eebf3e1)
- [OpenSourceJev Review: Logits, Not JSON](https://mrjev.com/projects/sabeel111-opensourcejev/): A local decision engine reading llama.cpp logits over a fixed candidate set. Linux needs two symlinks, and one model's calibration ships for all. (review, tested 2026-09-21 at 3c41fba)
- [OpenJev Review: One API Call Per Option](https://mrjev.com/projects/siliconlabai-openjev/): Not a model — a playground that scores every option with its own LLM call. Ignorance and certainty come out identical, and there's no licence. (review, tested 2026-09-21 at 443b679)
- [djev-run Review: A Cloud Run Recipe, Annotated](https://mrjev.com/projects/taeold-djev-run/): Deployment glue for serving DiffusionGemma-Jev on a Cloud Run GPU, with a zero-dependency Snake demo that says when it is only simulating. (review, tested 2026-09-21 at 9e74402)
- [LLM2Jev Review: Option Order Really Doesn't Matter](https://mrjev.com/projects/yinsongxu-llm2jev/): Turns a local model into typed decisions by scoring each candidate independently. We reversed the options and the probabilities were identical. (review, tested 2026-09-21 at 1909c50)
- [jev_local Review: A Real Drop-In, No Licence](https://mrjev.com/projects/argos1111-jev-local/): The only local project here the official SDK talks to unchanged. It has no LICENSE file, and its default weights bar companies over $10M. (review, tested 2026-09-20 at 8ccc04d)
- [jevmlx Review: Jev's Trick on Your Own Mac](https://mrjev.com/projects/bnsd55-jevmlx/): Rebuilds Jev's approach on local MLX models: one prefill, parallel rows, JSON assembled rather than parsed. Apple Silicon only, and the demo data is broken. (review, tested 2026-09-20 at 7ccf3ab)
- [OpenDecision Review: Jev's Shape, NLI Underneath](https://mrjev.com/projects/deepanwadhwa-opendecision/): Serves /v1/systemone from an off-the-shelf zero-shot classifier. Honest about what it is; ships with no authentication at all. (review, tested 2026-09-20 at 194dd11)
- [openJev-verdict-2.0: The Model Isn't There](https://mrjev.com/projects/heman10x-ngu-openjev-verdict-2-0/): Auditable metrics, a runnable training pipeline, checksummed artifacts — and the advertised checkpoint is a dead LFS pointer behind a private repo. (review, tested 2026-09-20 at a458733)
- [Bespoke Nimble Review: An Open 9B Decision Model](https://mrjev.com/projects/bespokelabsai-nimble/): Bespoke Labs' open recipe, data, and 9B model for Jev-style decisions. Carefully documented, but it needs a big GPU or Mac, and the code has no license. (review, updated 2026-09-19)
- [openjev-sglang Review: A Self-Hosted Jev-Style API](https://mrjev.com/projects/ekzhang-openjev-sglang/): A well-engineered server that speaks Jev's API on an open Qwen model. Works with TypeSafe's SDK, but needs a B200, has no license, and deploys unauthenticated. (review, tested 2026-09-19 at 604664a)
- [Simple Jev Review: Jev-Style Answers From Any Model](https://mrjev.com/projects/featherless-ai-simple-jev/): Featherless's server reads typed answers from open models' next-token scores, with a free demo API. It runs on a CPU but differs from Jev's API in several ways. (review, tested 2026-09-19 at 0dd5396)
- [LocalJev Review: Jev's API on Any Local Model](https://mrjev.com/projects/githubnext-localjev/): GitHub Next's small Bun server answers Jev requests with any OpenAI-compatible model. Highly compatible with Jev's API, but probabilities are self-reported. (review, tested 2026-09-19 at 3f23e36)
- [kev Review: Open Jev-Style Models From 0.5B to 8B](https://mrjev.com/projects/jaredpalmer-kev/): Jared Palmer's Jev-style decision models, with open weights, public training data, and a Jev-shaped API. The smallest runs on a CPU; a few API edges differ. (review, tested 2026-09-19 at 6f5ce16)
- [jeff Review: A Small Self-Hosted Jev-Style API](https://mrjev.com/projects/logan-markewich-jeff/): A self-hosted server that speaks Jev's API on a 400M-parameter GLiFormer model and runs on a plain CPU. TypeSafe's SDK works against it almost unchanged. (review, tested 2026-09-19 at 230d85d)
- [OpenJev Review: A Jev-Compatible Server on Diffusion](https://mrjev.com/projects/razorback16-openjev/): An Apache-2.0 decision server that speaks Jev's API on DiffusionGemma through vLLM, with image questions and a hosted option. It needs a large NVIDIA GPU. (review, updated 2026-09-19)
- [SemIf (OpenJev) Review: Jev-Style Decisions Locally](https://mrjev.com/projects/theoleecj-semif-openjev/): A rigorous open reproduction of Jev's interface on open models. Great for self-hosting and learning, but its scores are explicitly not calibrated confidence. (review, tested 2026-09-19 at b9cb325)
- [NanoJev Review: A 0.6B Open Jev Replica](https://mrjev.com/projects/tianyucodings-nanojev/): An ambitious, heavily documented 0.6B decision model with public weights and data. Needs a GPU, and some published training labels came from Jev itself. (review, updated 2026-09-19)
- [Jevlike Review: Train Your Own Jev-Style Scorer](https://mrjev.com/projects/vinnylarouge-jevlike/): A small, readable starter for training a one-pass option scorer on your own data. Trains in seconds on a CPU. A learning tool, not a Jev replacement. (review, tested 2026-09-19 at 94f5fd1)

## Reviews: Coding Agents & Developer Tools

- [mu Review: A Coding Agent With a Judgment Kernel](https://mrjev.com/projects/qybaihe-mu/): A pi-based coding agent that hands 38 routine calls to a judge, Jev by default. Built from source, its kernel tests pass, and the judge saw less than expected. (review, tested 2026-10-11 at faed8c3)
- [jev-rules Note: Only the Rules That Apply](https://mrjev.com/projects/eliaalberti-jev-rules/): A Claude Code plugin that shows Claude only the standing rules relevant to each prompt. Checked, not fully reviewed: tests, and what each prompt sends. (note, tested 2026-10-04 at a2b0dd3)
- [perch Review: A Linter That Reads the Call Graph](https://mrjev.com/projects/lakeday-org-perch/): It reads each method with its callers and callees, 29 questions in one request. Everything we found is fixed on main; two of them are not in a release yet. (review, tested 2026-09-20 at 54a38d6)
- [Jev Code Reviewer: One Read, No Writes](https://mrjev.com/projects/egma-ai-jev-code-reviewer/): Reviews your own agent's PRs on your machine. One GitHub read, no writes — and the diff and its file context go to both providers. (review, tested 2026-09-27 at d1966f5)
- [jevmem: Project Memory With a Careful State](https://mrjev.com/projects/avinash-jetwani-jevmem/): Automatic project memory for coding agents. Checked, not fully reviewed: its own suite is green and its request carries less than you would expect. (note, tested 2026-09-26 at b881810)
- [matchcn Review: It Pulled A Registry It Tagged](https://mrjev.com/projects/francesco0242-matchcn/): A semantic index over shadcn registries. Its limitations section explains why one registry was withdrawn after half its components failed to install. (review, tested 2026-09-22 at 2f61915)
- [jev-test-filter Review: It Fails Towards More](https://mrjev.com/projects/mizchi-jev-test-filter/): Picks which tests a diff could break and hands the arguments to your runner. Every failure we could cause made it run the whole suite instead. (review, tested 2026-09-22 at 01319cd)
- [patdown Review: Prose Rules, and It Runs Now](https://mrjev.com/projects/tyler-dot-earth-patdown/): Write a coding rule in Markdown and have a model check every file against it. The install and the credential skip list were both fixed in a day. (review, tested 2026-09-22 at 31810ef)
- [pi-jev Review: A Gate That Fails Open On Purpose](https://mrjev.com/projects/y0usaf-pi-jev/): A Jev gate for the Pi coding agent that judges bash, write and edit in one request — and lets all of them through when the provider breaks. (review, tested 2026-09-22 at 80a80f3)
- [JevLoop Review: The Model Picks, Code Offers](https://mrjev.com/projects/zjunlp-jevloop/): An agent loop where every fork is a typed decision and only writing costs an LLM call. The demo runs offline with no key and nothing installed. (review, tested 2026-09-22 at e67318c)
- [Skill Router Review: Typed Routing That Abstains](https://mrjev.com/projects/godsboy-jev-agent-skill-router/): Routes a request to one of 24 skills, or to review. Both outcomes reproduced against our stand-in, and the evaluation protocol is the best part. (review, tested 2026-09-21 at e481a06)
- [pi-jev-router Review: Asks Before the Prompt Lands](https://mrjev.com/projects/sugarforever-yummy-pi-extensions/): A per-turn session router for pi: continue, side-chat, fork or start fresh. A prefilter settles most prompts before any call is made. (review, tested 2026-09-21 at 0c19d05)
- [oxlint-plugin-jev: Lint Rules in Plain English](https://mrjev.com/projects/wobsoriano-oxlint-plugin-jev/): Write a lint rule as a question and oxlint enforces it. One request per file, a real cache, and a cap that quietly stops at 25 matches. (review, tested 2026-09-21 at 82703b8)
- [Jev Codex Router Review: Per-Turn Codex Routing](https://mrjev.com/projects/0xnatoshi-jev-codex-router/): A local proxy that lets Jev pick the model, reasoning depth, and speed for every Codex call. It works as documented, but can send work to DeepSeek and GLM. (review, tested 2026-09-19 at 8292b51)
- [Stanley Review: Bounded Workflows for Coding Agents](https://mrjev.com/projects/devagrawal09-stanley-code/): Careful evidence discipline, redaction and budgets — and a delegation path that runs an agent with approvals off on text a stranger wrote. (review, tested 2026-09-20 at 85f39e7)
- [winnow Review: Hiding Tool Output You Won't Read](https://mrjev.com/projects/ghalebdweikat-winnow/): Judges each block of a tool result and hides what the agent won't need, with honest calibration data. Its local service accepts anyone's requests. (review, tested 2026-09-20 at 51d80b9)
- [JevLint Review: Lint Rules in Plain English](https://mrjev.com/projects/iamtoomas-jevlint/): Write a coding standard as a sentence and it judges each file against it. Its credential exclusion missed half the extensions it scans; now fixed. (review, tested 2026-09-20 at 96b9d69)
- [compact-adviser Review: Knowing When to /compact](https://mrjev.com/projects/kunchenguid-compact-adviser/): Asks Jev whether your agent session is at a safe point to compact. The judgment is sharp, and it stays quiet far more often than it speaks. (review, tested 2026-09-20 at b2a27b5)
- [jev-lint Review: Two Walkers, One Disagreement](https://mrjev.com/projects/mizchi-jev-lint/): Checks whether a function does what its name says, with a cutoff per rule. Half the run respects .gitignore and the other half doesn't. (review, tested 2026-09-20 at 5f473f1)
- [agent-dispatcher Review: Closes the Bypass](https://mrjev.com/projects/nahid-sparktales-agent-dispatcher/): 27 agent roles with an optional Jev router, all scopes off by default. The most careful codebase here; its redactor misses AWS_SECRET_ACCESS_KEY. (review, tested 2026-09-20 at a709161)
- [jot Review: Every Word Is a Jev Decision](https://mrjev.com/projects/runta-dev-jot/): A local chat app where the tool, the arguments and each word of the reply are all Jev choices. One tool skips the sign-in gate. (review, tested 2026-09-20 at 55c3082)
- [jev-pruner Review: Trimming Bash Output](https://mrjev.com/projects/tamaratran-jev-pruner/): Asks Jev which lines of a long command's output the agent still needs, and drops the rest. Careful about losing things; not careful about sending them. (review, tested 2026-09-20 at 47d017c)
- [Jev Review: A Staged Code Reviewer Built on Jev](https://mrjev.com/projects/devagrawal09-jev-review/): A tidy, Jev-only code-review pipeline for JavaScript and TypeScript with a local dashboard. Good to learn from; it flags concerns, it doesn't explain them. (review, tested 2026-09-19 at 31f8960)
- [Jev Review MCP Review: Quality Scores for Agents](https://mrjev.com/projects/niazmorshed2007-jev-review/): A tidy MCP server that gives coding agents per-dimension quality scores from one Jev call. A useful signal, but it's built to send your diffs on every task. (review, tested 2026-09-19 at 57690af)
- [fast-jev-compaction Review: Claude Code Compaction](https://mrjev.com/projects/tamaratran-fast-jev-compaction/): A clever plugin that keeps Claude Code history verbatim instead of summarizing it. Promising but early: code you write goes to TypeSafe, and bugs pile up. (review, tested 2026-09-19 at e3f262a)
- [Foreman Review: Jev Supervising Codex Workers](https://mrjev.com/projects/thruwire-foreman/): A careful experiment that puts Jev in charge of watching a Codex coding agent. Thoughtful policy and limits, but diffs go to TypeSafe and Codex runs unattended. (review, tested 2026-09-19 at 64388e7)

## Reviews: Apps & Browser Extensions

- [Varina Review: Many Minds Diverge, Jev Prunes](https://mrjev.com/projects/deillusion-aha-engine/): A design-exploration engine: several LLM seats propose ideas and Jev prunes the duplicates. Its Jev layer is careful and fails safe; the repo ships no tests. (review, tested 2026-10-10 at 6ab3d23)
- [BridgeClip Note: Jev Checks Each Cut Holds Up](https://mrjev.com/projects/bridge-mind-bridgeclip/): A desktop video clipper that can ask Jev, via OpenRouter, whether each clip stands on its own. Checked, not fully reviewed: a very careful Jev client. (note, tested 2026-10-07 at b996820)
- [Jevbox Review: Permission-Aware Search, Kept Tight](https://mrjev.com/projects/extend-hq-jevbox/): A self-hosted document library where Jev walks the folder tree and scores passages. Its permission checks held under our probe; one 529 fails a search. (review, tested 2026-10-07 at aaa9381)
- [OpenMuse Review: Jev Decides How to Ask, Not What](https://mrjev.com/projects/copilotkit-openmuse/): CopilotKit's self-hosted personal agent uses Jev for one narrow job, done carefully. Running it needs CopilotKit's hosted service, telemetry on by default. (review, tested 2026-10-06 at 73a7149)
- [dasheng: Pronunciation Scoring, Judged Typed](https://mrjev.com/projects/wquguru-dasheng/): Read English aloud and see which words were wrong. Checked, not fully reviewed: the judging half runs against a stand-in, the ASR half needs a GPU. (note, tested 2026-09-27 at 1bacff4)
- [Shapeshift Review: One Box, Fourteen Questions](https://mrjev.com/projects/anishfn-shapeshift/): A text box that becomes the right small interface. One request per keystroke-burst carries only what you typed, and it stays offline until a real key appears. (review, tested 2026-09-26 at 5e24166)
- [Passage Review: It Leads With Its Own Failures](https://mrjev.com/projects/austinaway-working-memory-jev/): A reading tool that flags where a passage may overload working memory — and whose evaluation doc opens by naming the two tests its model fails. (review, tested 2026-09-22 at cee3edb)
- [Jev Demos Review: Your Key Cannot Echo Back](https://mrjev.com/projects/mayank953-jev/): Seven side-by-side demos that run with no keys at all, label themselves simulated, and redact your key out of any upstream error. (review, tested 2026-09-22 at 4d5d69f)
- [Cheshi Review: The Key No Longer Does Two Jobs](https://mrjev.com/projects/cheshiai-cheshi/): A macOS workspace for Codex with Jev-powered conversation recall. Saving a key used to enable it silently; since build 9 it has its own switch, off by default. (review, tested 2026-09-20 at 80b61db)
- [Jev Explained Review: A Playground With Receipts](https://mrjev.com/projects/davila7-jev-explained/): An interactive explainer with four worked examples and a request trace. Your key rides in a header through its own server on every call. (review, tested 2026-09-21 at 5cbe35e)
- [Crush Monitor Review: Reading a Chat, Carefully](https://mrjev.com/projects/ferrycorleone-crush-monitor/): Paste a WeChat conversation and it labels emotion, intent and how your replies landed. It told us our sample was insufficient, which is the right answer. (review, tested 2026-09-21 at be1b33b)
- [Jeved Review: Rules for a SillyTavern Story](https://mrjev.com/projects/mossyfield-st-jeved/): Writes your own questions about each reply and acts on the answers: nudge the narrator, reroll, or run a script. 621 tests for a two-day-old extension. (review, tested 2026-09-21 at 040566c)
- [jev-mail Review: Inbox Actions, One Config Bug](https://mrjev.com/projects/parth-kp-jev-mail-classifier/): Tags, moves and flags real mail from one noul per category. We ran it against a real IMAP server, and the shipped webhook config stops the run. (review, tested 2026-09-21 at db96194)
- [Vibe Check Review: Score a Post Before You Send It](https://mrjev.com/projects/rafalwilinski-vibecheck/): A Chrome extension that scores your draft on a dozen rubrics before you hit Post. Minimal permissions, but the image toggle is on before you add a key. (review, tested 2026-09-21 at d834d50)
- [jev-paint Review: Painting With Probabilities](https://mrjev.com/projects/achimala-jev-paint/): Asks Jev for the colour of every pixel and renders the distributions as a painting. A small local experiment where every README claim checked out. (review, tested 2026-09-20 at ecf9c48)
- [Jev Moderation Bot Review: A Discord Mod on Rails](https://mrjev.com/projects/brainstormity-jev-moderation-bot/): Classifies every message and escalates through a warning ladder code controls. When the API fails, everything is let through, silently. (review, tested 2026-09-20 at 1629ac8)
- [jevmeter Review: A Live Meter Over a Debate](https://mrjev.com/projects/chetaslua-jevmeter/): Scores every sentence of a video and renders a 16:9 edit with live probability bars. Its cache can quietly reuse old scores after you edit the transcript. (review, tested 2026-09-20 at cbf8e11)
- [jev-reviewer Review: Anything After the References](https://mrjev.com/projects/choxos-jev-reviewer/): Extracts trial data as verbatim quoted lines, never generated text. Content after the reference list is dropped and exported as 'Not reported'. (review, tested 2026-09-20 at da15868)
- [Lurk Review: Finding Buyers on Reddit With Jev](https://mrjev.com/projects/getanyapi-com-lurk/): A self-hostable tool that judges Reddit posts for buying intent in three graded passes. Well engineered, and it sends strangers' words to a third party. (review, tested 2026-09-20 at 18665a8)
- [live-jev Review: Voice Control for Ableton Live](https://mrjev.com/projects/okinaaudio-live-jev/): Say a phrase in Japanese or English and it moves a fader. A local phrase table catches most of it; what misses sends your whole session. (review, tested 2026-09-20 at ecba7e0)
- [x-scanner Review: Honest About Everything But One](https://mrjev.com/projects/oso95-x-scanner/): Judges every X post you scroll past on six axes. Its privacy policy is unusually accurate, and it never mentions protected accounts. (review, tested 2026-09-20 at fd600dd)
- [typesafe-adblock Review: Judging Ads by Meaning](https://mrjev.com/projects/realzachi-typesafe-adblock/): A Chrome extension that asks Jev whether an element is an ad, then removes it. Clever, tiny, and it removes full-screen iframes without the size guard. (review, tested 2026-09-20 at 7e067d2)
- [sharp Review: An X Timeline Filter That Reads](https://mrjev.com/projects/tshmieldev-sharp/): Hides posts by a rule you write in your own words, with local filters doing the cheap work first. The auth-header finding was fixed the same day. (review, tested 2026-09-20 at 6afca0c)
- [jev-skip Review: Sponsor Skipping Without a Crowd](https://mrjev.com/projects/valentynkit-jev-skip/): Finds YouTube sponsor segments on videos nobody has labelled, by judging transcript windows. Honest about what it catches and what it costs you. (review, tested 2026-09-20 at 6837e3e)
- [jev-chat Review: A Chat Assistant With No LLM](https://mrjev.com/projects/w3cj-jev-chat/): Every reply is template text or tool data; Jev only picks among options code supplies. Two destructive actions skip the confirmation step. (review, tested 2026-09-20 at e543aba)
- [Unclutter Review: A Jev Ad and Clutter Hider](https://mrjev.com/projects/kitze-unclutter/): A carefully designed browser extension that learns reusable hiding rules per page type. Its privacy claims held up in our tests; you install it from source. (review, tested 2026-09-19 at 9ef9bec)

## Reviews: Agent Integrations (MCP & Skills)

- [Quicksilver Note: Hand Claude's Skimming to Jev](https://mrjev.com/projects/uditakhourii-quicksilver/): A Claude Code skill that hands bulk filter, classify and rank calls to Jev. Checked, not fully reviewed: skips secret-named files, not passwords in config. (note, tested 2026-10-10 at 5d6fe5c)
- [Gatekeeper Note: Rules First, Then One Jev Call](https://mrjev.com/projects/agricidaniel-gatekeeper/): Routes requests to an agent or skill with YAML rules plus one Jev choice. Checked, not fully reviewed: redacts common keys, fails open, starts in shadow mode. (note, tested 2026-10-06 at bcc1c20)
- [System One Connector Review: One Tool, Many Calls](https://mrjev.com/projects/itsmostafa-system-one-connector/): One read-only MCP tool that gives an agent typed decisions, with careful JSON handling — and a batch mode where one approved call becomes 500 billed requests. (review, tested 2026-09-28 at 2124fe6)
- [dsh-jev-interceptor: A Guard That Delegates](https://mrjev.com/projects/asktheway-dsh-jev-interceptor/): Judges every tool call in DeepSeek Harness. Checked, not fully reviewed: 64 tests green, and its state is three bounded fields by design. (note, tested 2026-09-26 at ee8f4af)
- [Hermes Approvals Review: Six Questions, One Ladder](https://mrjev.com/projects/anpicasso-hermes-jev-approvals/): Replaces Hermes Agent's approval reviewer with six typed questions and deterministic code. We walked every branch of the ladder ourselves. (review, tested 2026-09-22 at 28be98a)
- [Skill Router Review: Two Gates, Then Silence](https://mrjev.com/projects/decrux9812-typesafe-skill-router/): A Hermes plugin that names the one skill worth loading. When the first gate is not cleared it suggests nothing and does not spend the second request. (review, tested 2026-09-22 at 94fe114)
- [Jevify Review: A Skill That Teaches the Habit](https://mrjev.com/projects/altryne-jevify/): A skill that tells your agent when to reach for Jev, with a stdlib client and a document scanner that counts what it failed to judge. (review, tested 2026-09-21 at 787f9a5)
- [Jev DSH Review: A Decision Plugin for DeepSeek](https://mrjev.com/projects/devin-axis-jev-dsh-decision/): A 200-line decision service for DeepSeek Harness and iPolloWork. We fed it failing providers with the key in the error, and nothing leaked. (review, tested 2026-09-21 at adc88ca)
- [lorenzini Review: Never Infer a Pass From Silence](https://mrjev.com/projects/nanako0129-lorenzini/): Three skills that wait for CodeRabbit, Copilot and Codex and decide whether the verdict really permits a merge. Jev is the shadow, not the gate. (review, tested 2026-09-21 at 06c79ea)
- [JCR Review: One Tool Instead of Ten Skill Files](https://mrjev.com/projects/niazmorshed2007-jcr/): A resolver that walks 11,360 documented operations with Jev and hands your agent only the ones it picked. We drove the walk with a stand-in. (review, tested 2026-09-21 at 138b383)
- [Jevvy Review: Auto-Approval That Fails Closed](https://mrjev.com/projects/panachy-jevvy/): Four typed questions decide whether a shell command needs you. Every failure path we could force ended in "ask", and it sends only the command. (review, tested 2026-09-21 at 90f2736)
- [jev-use Review: A Gate That Only Tightens](https://mrjev.com/projects/shitianfang-jev-use/): MCP tools and a PreToolUse gate for Claude Code, Codex and pi. Every outcome we forced — deny, ask, silence, outage — landed where it should. (review, tested 2026-09-21 at 358819d)
- [pi-jev Review: Tool Routing for the Pi Agent](https://mrjev.com/projects/theooliveira-pi-jev/): Semantic tool routing for the Pi agent. Its failure path used to switch on the first three tools it found; since 549c2bf it switches on none. (review, tested 2026-09-20 at 0c3ac04)
- [Jev Studio Review: A Kit With a Broken Front Door](https://mrjev.com/projects/utk2103-jev-studio/): A Python port of jev-cli plus prompt libraries and an MCP server. The MCP server won't start on a fresh install, and the port inherited two fixed bugs. (review, tested 2026-09-21 at fe47b78)
- [Jev Skills Review: Six Skills, One Careful CLI](https://mrjev.com/projects/wuyoscar-jev-skill/): A skills collection for coding agents whose 300-line client is the most careful Jev caller we've read: allow-listed endpoints and no redirects. (review, tested 2026-09-21 at 36b08bb)
- [jevwire Review: One Copy of the Rule, With a Typo](https://mrjev.com/projects/brainwires-jevwire/): The most thorough guardrail here: root-confined reads, priced before sending, 1,461 tests. One entry in its sensitive-directory set can never match. (review, tested 2026-09-20 at fabe7e7)
- [jev-mcp (burnigtm): Redaction That Blinds](https://mrjev.com/projects/burnigtm-jev-mcp/): Nine tools, a token budget, and the only real redaction in its batch: it sends argument shapes, not values. So the model approves calls it never sees. (review, tested 2026-09-20 at 4f4ae11)
- [jev-eval-mcp Review: Pick a Threshold From Data](https://mrjev.com/projects/byk-jev-mcp/): An MCP server whose point is the sweep: run a labelled set, see precision and recall at every threshold. One tool writes wherever it's told. (review, tested 2026-09-20 at cee2e6d)
- [jev-sift Review: Read It Without Reading It](https://mrjev.com/projects/kbhuw-jev-sift/): An MCP server that classifies files and pages so your agent never loads them. Careful SSRF work, no licence, and its default root is your home directory. (review, tested 2026-09-20 at 966de12)
- [Hermes Jev Skills: Eight Small Decisions](https://mrjev.com/projects/kerpopule-hermes-jev-skills/): Routing, skill selection, retrieval filtering and compaction, each handed to Jev. Everything we reported was fixed in 0.18.0 within hours. (review, tested 2026-09-20 at a5d8901)
- [askjev Review: Let the Model Pick the Question Type](https://mrjev.com/projects/pzacca-askjev/): One MCP tool: write a question in plain words and it routes to yes/no, choice or score first, then answers it. Clean errors, no stars yet. (review, tested 2026-09-20 at 77b6d79)
- [Jevbridge Review: The Gate That Didn't Gate](https://mrjev.com/projects/tacticocc-jevbridge/): An MCP and ACP server that batches many questions into one call. Its gate was broken two ways when we reviewed it; both were fixed the same day. (review, tested 2026-09-20 at 54c5587)
- [Building with Jev Skill Review: Agent Guidance](https://mrjev.com/projects/dbreunig-building-with-jev-skill/): A concise, accurate agent skill for designing Jev questions and fixing ones that answer wrong. Every claim we checked matched TypeSafe's docs. No license yet. (review, tested 2026-09-19 at 04fe366)
- [Jev MCP Review: Ten Judgment Tools for Agents](https://mrjev.com/projects/jkudish-jev-mcp/): A well-built MCP server that gives agents ten focused Jev tools, from fact-checking to extraction. All ten worked in our tests, with careful limits throughout. (review, tested 2026-09-19 at 67dd9fa)

## Reviews: Libraries & Integrations

- [JEV Sees Note: Camera Facts In, Jev Answers Out](https://mrjev.com/projects/charlesfeng0314-jev-sees/): A local vision front end that turns frames into text facts for Jev, so no pixels leave. Checked, not fully reviewed: sound design; installs need a pin. (note, tested 2026-10-11 at 8cf89cb)
- [Ten Levels of Jev Note: Thirty Worked Uses](https://mrjev.com/projects/disler-ten-levels-of-jev/): Thirty concrete uses of Jev, from an if-statement to an agent that reaches for it itself. Checked, not fully reviewed: its suite and its offline demo. (note, tested 2026-10-04 at 777adaf)
- [Discern Note: Semantic Control Flow for Effect](https://mrjev.com/projects/doeixd-discern/): Patterns, policies and procedures over Effect's Decision and DecisionModel, where 'maybe' is a branch. Checked, not fully reviewed: its suite. (note, tested 2026-10-04 at 95c6791)
- [system-one Review: One Decision, Any Provider](https://mrjev.com/projects/iamaamir-system-one/): A TypeScript runtime that keeps decision code the same across providers and rejects malformed answers rather than smoothing them over. Our one gap is fixed. (review, tested 2026-09-29 at 921d86d)
- [Jev Cookbook Note: Fifteen Recipes With Receipts](https://mrjev.com/projects/nexibeo-jev-cookbook/): Fifteen runnable recipes, each shipping the output of the run that produced it. Checked, not fully reviewed: what the recipes record, not their answers. (note, tested 2026-10-04 at 3e0207d)
- [pijev Note: Averaging Out Option Order](https://mrjev.com/projects/typellm-pijev/): Asks each Choice in a balanced set of option orders in one request and averages the answers. Checked, not fully reviewed: the averaging is exact. (note, tested 2026-10-04 at c85474a)
- [Truffler Review: Search Your Rails App by Intent](https://mrjev.com/projects/kieranklaassen-truffler/): Labels every record with typed questions once, then ranks in SQL. Searching sends the query and no record content, and the read-list really holds. (review, tested 2026-10-01 at 87e373c)
- [jev-feels Note: A Question as a Predicate](https://mrjev.com/projects/qew7-jev-feels/): email.feels?(:urgent) as a named question on a field, with an Active Model validator. Checked, not fully reviewed: its suite, and what each call sends. (note, tested 2026-10-01 at 76fa9af)
- [typesafe.pro Review: Someone Else's Key](https://mrjev.com/projects/chigwell-typesafe-pro/): A free anonymous endpoint that speaks Jev's API. We ran the proxy ourselves: your state goes to TypeSafe on the operator's account, verbatim. (review, tested 2026-09-22 at 360e648)
- [feelings Review: The AI If-Statement, Typed](https://mrjev.com/projects/boundaryml-feelings/): BAML's `.feels()` on any value: Jev answers the judgment, an LLM writes the prose, and the return type is the schema it must fill. (review, tested 2026-09-21 at e0f07b4)
- [Hunch Review: Probabilistic If-Statements in Ruby](https://mrjev.com/projects/carldaws-hunch/): chance, pick and rate as Ruby methods, with a stub backend for tests and one API call per decide block. It reads like Ruby, not like an SDK. (review, tested 2026-09-21 at 4651c3e)
- [System One Harness: An Agent Loop That Picks](https://mrjev.com/projects/harnessrouter-systemoneharness/): Compiles an environment's actions into typed questions, then gates each one by risk: read at 0.5, write at 0.7, destructive at 0.9. (review, tested 2026-09-21 at 5768c4b)
- [Jevalyn Review: A Decision Layer for Rails](https://mrjev.com/projects/ray-hughes-jevalyn/): Decisions as classes, validated when the app boots, with a confidence threshold and a mock mode that never opens a connection. (review, tested 2026-09-21 at 0567d10)
- [HA-Jev Review: Jev in Home Assistant](https://mrjev.com/projects/abovecolin-ha-jev/): A carefully built integration that answers house questions with typed decisions. We found its locks inverted; the author removed locks entirely in 1.9.1. (review, tested 2026-09-20 at 1f63190)
- [jev-harness Review: The Parts Around the Answer](https://mrjev.com/projects/antoniocoppe-jev-harness/): A small TypeScript library for the glue nobody enjoys writing: a policy, a confidence gate, shadow mode, reusable recipes and an offline eval CLI. (review, tested 2026-09-20 at 2934d12)
- [invalidate Review: Memories With an Expiry Vote](https://mrjev.com/projects/chopratejas-invalidate/): Stores agent memories with leases and asks Jev whether a new event kills them. Careful work, but it is not on PyPI and under-counts its own requests. (review, tested 2026-09-20 at d6ade60)
- [jevcache Review: Bob Gets Alice's Decision](https://mrjev.com/projects/hyperspaceai-jevcache/): A local cache for Jev decisions, shipped as a binary with no source and no licence. It drops user_id before hashing, so two users share one answer. (review, tested 2026-09-20 at v0.1.0)
- [zod-jev Review: Zod Checks Shape, Jev Meaning](https://mrjev.com/projects/jomatsu-zod-jev/): Adds semantic rules to a Zod 4 schema, sends them all in one request, and turns probabilities into ordinary Zod issues. Neatly done. (review, tested 2026-09-20 at 700bd25)
- [ruby_llm-typesafe Review: Jev as a RubyLLM Provider](https://mrjev.com/projects/kieranklaassen-ruby-llm-typesafe/): A clean provider that registers Jev with RubyLLM and leaves the thresholds to you. Two commits old, and already behind the gem it plugs into. (review, tested 2026-09-20 at 33e6801)
- [jev-tree Review: Choosing From Thousands of Options](https://mrjev.com/projects/reachjalil-jev-tree/): A Choice tops out at 255 options; real catalogues are bigger. This walks a taxonomy and asks once per level, so a deep tree becomes a few calls. (review, tested 2026-09-20 at 95bff63)
- [hono-jev-router Review: Routing by Description](https://mrjev.com/projects/yusukebe-hono-jev-router/): Register a Hono handler with a sentence instead of a path. A neat experiment, but a Jev failure takes the whole app down with it. (review, tested 2026-09-20 at 04f6e10)
- [n8n-nodes-typesafe: Typed Questions in a Workflow](https://mrjev.com/projects/zampierid4p-n8n-nodes-typesafe/): Fills in the form, sends one request per item, and can fan the answers out to one branch per question. No tests, and a retry that can sleep for hours. (review, tested 2026-09-20 at 98bde6a)
- [Advocaat Review: A Tiny Typed TypeScript Client](https://mrjev.com/projects/pithings-advocaat/): A small, dependency-free TypeScript client that makes Jev questions read like code. Everything we tried matched its README. Early, but carefully built. (review, tested 2026-09-19 at bc46287)

## Reviews: Command-Line Tools

- [JevRev Review: Careful Audits, Trusted Paperwork](https://mrjev.com/projects/alex314618-create-jevrev/): A mostly deterministic LLM-plus-Jev loop that keeps your logs at home. But a hand-written evidence file completes it, and Jev grades summaries, not work. (review, tested 2026-10-05 at 38d83f3)
- [jym Note: Did You Mean, With a Safety Catch](https://mrjev.com/projects/syumai-jevyoumean/): Fixes mistyped subcommands with one Jev choice question. Checked, not fully reviewed: the denylist holds in auto mode, and your arguments stay local. (note, tested 2026-10-05 at f0bed70)
- [jevgrep Review: Ask Where the Code Is](https://mrjev.com/projects/dzhng-jevgrep/): A retrieval CLI for coding agents with unusually strict evidence for its cost claim — and, on its own numbers, a saving carried by two tasks out of ten. (review, tested 2026-09-29 at 82ef1fd)
- [webctl Review: A Benchmark That Excludes Itself](https://mrjev.com/projects/dorkitude-webctl/): A search CLI that scores results with Jev so your agent reads less, with the most careful benchmark write-up in this directory. (review, tested 2026-09-23 at e9bc54a)
- [jgrep Review: Grep's Exit Codes, Mostly Kept](https://mrjev.com/projects/kyu1204-jgrep/): Semantic grep that packs sixteen chunks and sixteen questions per request. We counted the packing, and every failure we could force now exits 2. (review, tested 2026-09-23 at 9961c60)
- [slop-grader Review: Counted On The Wire](https://mrjev.com/projects/lukstei-slop-grader/): Grades writing line by line against 21 rules. Every rule really is asked about every line, and a short answer set is now an error rather than a clean report. (review, tested 2026-09-23 at db60009)
- [jev-cli Review: The Key Stays Out Of argv](https://mrjev.com/projects/tumf-jev-cli/): A CLI and stdio MCP server for the three Jev primitives, with structured errors and real exit codes on every failure path we could force. (review, tested 2026-09-23 at 980cb98)
- [evoke Review: A Sentence, Gated Into a Call](https://mrjev.com/projects/evoke-build-evoke/): A sentence becomes a call of a small program you installed, gated on confidence. Every 'this may only tighten' rule we could reach, we tried to loosen. (review, tested 2026-09-22 at 6d74482)
- [jev (CLI) Review: An Exit Code For Not Sure](https://mrjev.com/projects/shaharia-lab-jev-cli/): A Rust CLI whose exit codes distinguish no from the model could not tell — and which lints your question before it spends anything. (review, tested 2026-09-22 at 8b048a6)
- [jegrep Review: Three Passes, No Index](https://mrjev.com/projects/can1357-jegrep/): A cascade that narrows by filename, then a 384-byte sketch, then the real text. It learned what a credential file is the day we reported it. (review, tested 2026-09-20 at a280f14)
- [jev-cli Review: The Best-Built One Here](https://mrjev.com/projects/nasrallah-al-jev-cli/): Thirteen Jev subcommands with CI-shaped exit codes, real release hygiene and the best key handling here. The flag gaps we found shipped fixed in 0.2.3. (review, tested 2026-09-21 at b4eae54)
- [JevGrep Review: Search a Repo by Behaviour](https://mrjev.com/projects/nassim-arifette-jevgrep/): Ask what the code does and get the excerpt back. Remote evaluation is off until you turn it on, and inspect shows the bill first. (review, tested 2026-09-21 at 52e2d89)
- [jeff Review: Written Rules, Checked in CI](https://mrjev.com/projects/alurith-jeff/): A Go CLI that checks files against twenty written rules with a three-way verdict. It stopped uploading keys and configs the day after we reported it. (review, tested 2026-09-20 at 656ca43)
- [jev-repl Review: A REPL for Shaping Questions](https://mrjev.com/projects/aoprisan-jev-ts-repl/): Write a question in one line, see the bars, then take the request to code. Zero dependencies, 157 tests, and the most careful key handling in this batch. (review, tested 2026-09-20 at cfabe94)
- [Sniff Test Review: Consent Before Anything Leaves](https://mrjev.com/projects/danrwilloughby-snifftest/): A prose linter for AI writing tells, with the most careful consent gate we've seen. Name a file explicitly and the extension check is skipped. (review, tested 2026-09-20 at 240653e)
- [commit-miner Review: Classifying Git History](https://mrjev.com/projects/devanshbatham-commit-miner/): Rust that asks Jev what each commit actually did — bug fix, security fix, change type — with genuinely hardened Git handling. Raw diffs leave unredacted. (review, tested 2026-09-20 at 977617e)
- [jgrep Review: Grep Where the Pattern Is a Sentence](https://mrjev.com/projects/keltokhy-jgrep/): Filters lines by meaning, for pipes as well as files, with a dollar budget and a local estimate first. Four of its six skip rules only apply with -r. (review, tested 2026-09-20 at bc0dc8d)
- [triagedy Review: Alert Triage as a Unix Filter](https://mrjev.com/projects/m0rphtail-triagedy/): JSONL alerts in, typed decisions out, with the routing done in Rust. The whole alert goes out unredacted, and one confidence is never checked. (review, tested 2026-09-20 at 5bef4e6)
- [typesafe-ai-playground: Five Experiments in Rust](https://mrjev.com/projects/markjaquith-typesafe-ai-playground/): PHI detection, comment quality, a load-bearing heat map, tone and classification. Careful about prompt injection, careless about how many calls it makes. (review, tested 2026-09-20 at 344d774)
- [jev-shell-history Review: Jev at Your zsh Prompt](https://mrjev.com/projects/mrnugget-jev-shell-history/): Fish-style zsh autosuggestions ranked by Jev, including plain-English ones. It works well, but it sends your recent shell history on every keystroke. (review, tested 2026-09-20 at 4b2b75d)
- [SemDecide Review: Judgment for Unix Pipelines](https://mrjev.com/projects/sharziki-semdecide/): grep for meaning and jq for judgment, with exit codes a CI job can branch on. Small, dependency-free, and it does the boring parts for you. (review, tested 2026-09-20 at 33cf5c0)
- [sys1grep Review: Grep by Meaning, in Any Language](https://mrjev.com/projects/uehaj-sys1grep/): One file, no dependencies, and a grep that matches a proposition rather than a pattern. Every claim in its README reproduced on our run. (review, tested 2026-09-20 at ba6ef50)
- [jev-commit Review: Does the Message Match the Diff?](https://mrjev.com/projects/valentynkit-jev-commit/): A commit-msg hook that checks your message against the staged diff, and blocks credential-shaped lines — after it has already sent them. (review, tested 2026-09-20 at 311e163)

## Reviews: Browser & Computer Use

- [dejevu Note: A Browser Agent That Skips Jev](https://mrjev.com/projects/idovmamane-dejevu/): A one-look browser agent on any open model; Jev is an optional comparison arm. Checked: that backend behaves well, and the head-to-head is not a rerun. (note, tested 2026-10-05 at 21df0bd)
- [Jev Browser Review: An Index, Not a Screenshot](https://mrjev.com/projects/openqa-cn-jev-browser/): Browser automation that indexes the page's own controls and sends no pixels. Password values stay behind; the plan still needs a second, generative model. (review, tested 2026-09-26 at d839d10)
- [fastbrowse Review: Answers With Receipts](https://mrjev.com/projects/agent-labs-dev-fastbrowse/): Jev picks each action from what is on the page, an LLM reads, and every claim in the answer carries a verbatim quote with a hash of the capture. (review, tested 2026-09-22 at ce08a6c)
- [Jev Browser Use Review: DONE Never Means Done](https://mrjev.com/projects/wy-coliney-jev-browser-use/): A Codex browser skill where Jev picks one action from a list code built, and its strongest answer — DONE — only hands control back for a check. (review, tested 2026-09-22 at f14b60e)
- [Jev for Chrome: The Tab You Are Looking At](https://mrjev.com/projects/chy4pro-jev-for-chrome/): An MV3 port of jev-ultrafast that drives your current tab. Password values stay out of the snapshot, and Enter is a choice the model has to make. (review, tested 2026-09-21 at a6ddd3a)
- [jev-browser Review: Headless Browsing With Jev](https://mrjev.com/projects/jkudish-jev-browser/): A well-documented headless browser agent with an auditable trace. Worked on simple pages, but typing presses Enter, which can submit forms you meant to fill. (review, tested 2026-09-19 at 8d90c51)
- [JevScout Review: A Job Hunter in Your Browser](https://mrjev.com/projects/hqman-jevscout/): Drives your logged-in Chrome to find job postings, scoring links with Jev. One ATS redirect empties the report, and it can type into focused fields. (review, tested 2026-09-20 at b51390b)
- [jev-macos-loop Review: Checks Its Own Homework](https://mrjev.com/projects/jcpsimmons-jev-macos-loop/): Native macOS computer use where Jev picks each action from text only. The one project here that verifies the outcome instead of trusting DONE. (review, tested 2026-09-20 at 1aadc01)
- [Jev Voice Review: The Wake Word That Isn't](https://mrjev.com/projects/kevinbadi-jev-voice/): Talk to your Mac and one request turns the transcript into an action. Follow the install instructions and every sentence in the room is sent, named or not. (review, tested 2026-09-20 at cc69795)
- [voice-browser Review: Talking to a Real Browser](https://mrjev.com/projects/moritzkremb-jev-voice-browser/): Speak, and a real Chromium acts before you finish the sentence. It worked on every step we tried — but it sends the contents of filled form fields. (review, tested 2026-09-20 at 054db0f)
- [Jev-cu Review: Computer Use Without Screenshots](https://mrjev.com/projects/sac-y-jev-cu/): A Codex computer-use skill where Jev picks the element from on-screen text. The safety gates held in our tests, but one of them can be truncated away. (review, tested 2026-09-20 at 38fb31d)
- [jev-use Review: The Screen, Minus the Passwords](https://mrjev.com/projects/savka777-jev-use/): macOS computer use through the accessibility tree, with the cleanest key handling here. Everything on screen except secure fields goes out. (review, tested 2026-09-20 at 8907f85)
- [macbrow Review: A Policy Layer Worth Copying](https://mrjev.com/projects/timpratim-macbrow/): Voice to AppleScript, with three policy checkpoints and a spoken confirmation. Its own code-generation prompt recommends things the policy then blocks. (review, tested 2026-09-20 at a392a9c)
- [Jev Desktop Review: The One With a Budget](https://mrjev.com/projects/yikangy873-gif-jev-desktop/): A bounded decision loop inside Codex Computer Use, with the most restrained payload here. Its allow-list check accepts a regex matching everything. (review, tested 2026-09-20 at 9b02783)
- [typesafe-computer-use Review: Jev Drives a Mac](https://mrjev.com/projects/awlevin-typesafe-computer-use/): A careful macOS computer-use loop where Jev picks each action from OCR and accessibility data. Cheap per step, but screen text and a screenshot leave your Mac. (review, updated 2026-09-19)
- [jev-ultrafast Review: A Jev-Powered Browser Agent](https://mrjev.com/projects/browser-use-jev-ultrafast/): A fast, carefully engineered browser agent from Browser Use where Jev picks every action. Impressive, but by default it drives your real Chrome profile. (review, tested 2026-09-19 at 1231850)
- [Mobile Jev Review: Jev Drives an Android Phone](https://mrjev.com/projects/droidrun-mobile-jev/): A carefully built Android agent where Jev picks every tap and never invents text. Well tested, but it only works with DroidRun's paid Mobilerun cloud phones. (review, tested 2026-09-19 at 395fc22)
- [agent-desktop Review: Jev-Driven macOS Automation](https://mrjev.com/projects/lahfir-agent-desktop/): A mature macOS automation tool whose new jev-desktop loop lets Jev pick each click. Thoughtful design, but it's hard to find and needs broad permissions. (review, updated 2026-09-19)

## Reviews: Games & Real-Time Demos

- [is-jeven Note: Is It Even? Ask Jev](https://mrjev.com/projects/wobsoriano-is-jeven/): A joke package that asks Jev whether a number is even. Checked, not fully reviewed: a tidy example of one raw /v1/systemone call that says it is a joke. (note, tested 2026-10-07 at 0dd1960)
- [Jev experiments Review: 21 Demos, No Licence](https://mrjev.com/projects/dabit3-jev-experiments/): Twenty-one latency-focused Jev demos from Nader Dabit, built by Devin. Every web app's own tests pass; not one of them carries a licence. (review, tested 2026-09-22 at c469e5b)
- [Jev Tetris Review: A Fair Fight, No Licence](https://mrjev.com/projects/trungdq88-jev-tetris/): Two models play Tetris on the same seeded piece sequence under the same clock. The proxy that makes it work is 131 lines and stores nothing. (review, tested 2026-09-22 at a0b411c)
- [jevchat Review: A Chatbot Made of Yes/No](https://mrjev.com/projects/kyle-pena-nlp-jevchat/): Asks which symbol comes next and samples from the answer, turning a decision model into a text generator. It works, and it is one request per symbol. (review, tested 2026-09-21 at d58abd5)
- [jev-pong Review: The Ball Moves at Model Speed](https://mrjev.com/projects/ably-labs-jev-pong/): Pong where the paddle waits for a decision, so latency is something you watch rather than read. Its published numbers recompute from the repo. (review, tested 2026-09-20 at d28d6ae)
- [jev-chess-lab Review: An Honest Negative Result](https://mrjev.com/projects/denikuchero-jev-chess-lab/): One day of chess experiments, published whole: the deliberation loop cost eight times more per move and played worse. Written up in Russian. (review, tested 2026-09-20 at f501d9d)
- [1v1 Jev Review: Quickscoping Against a Model](https://mrjev.com/projects/emrickgarrett-onevonejev/): A browser FPS where the bot's movement and aim are typed choices, several a second — with a heuristic taking over when the API doesn't answer. (review, tested 2026-09-20 at 365b339)
- [tsai-sc Review: Jev Plays StarCraft, Paused](https://mrjev.com/projects/phyous-tsai-sc/): Reads game memory, asks Jev which command to issue, and won a shareware mission. Honest about almost everything — except who did the visual check. (review, tested 2026-09-20 at 6046ecc)
- [jev-drone Review: A Camera-Only Quadrotor](https://mrjev.com/projects/romanslack-jev-drone/): A simulated drone where Jev makes the tactical calls and code keeps it flying. The README's headline comparison no longer reproduces at HEAD. (review, tested 2026-09-20 at cbeb53c)
- [jev-plays-pokemon-red: Code Plans, Jev Picks](https://mrjev.com/projects/valentynkit-jev-plays-pokemon-red/): Python does the RAM decoding, pathfinding and type chart; Jev only chooses among options code already enumerated. The smallest payload we've measured. (review, tested 2026-09-20 at cbe5387)
- [live-jev Review: A Driving Sim That Hides Failures](https://mrjev.com/projects/vinilana-live-jev/): Jev steers a 2D car at five decisions a second. Every failed call silently falls back to local rules while the label still says Jev. (review, tested 2026-09-20 at cd13ab0)
- [TypeSafe Mario Review: Jev Plays Super Mario Bros.](https://mrjev.com/projects/fhshaik-typesafe-mario/): A well-structured demo where Jev picks Mario's controller input from parsed game state, not screenshots. Great to read; no license, and you need your own ROM. (review, updated 2026-09-19)
- [JevPilot Review: A Jev-Driven Driving Simulator](https://mrjev.com/projects/standardagents-jevpilot/): A striking Three.js driving demo where Jev picks the car's path several times a second. Great to play and study, but it has no license and some tests fail. (review, tested 2026-09-19 at e1beeb1)

## Reviews: Guardrails & Safety

- [dsh-jev Review: You Cannot Configure It To Allow](https://mrjev.com/projects/buberlo-dsh-jev/): A decision layer for DeepSeek Harness whose failure policy refuses the value 'allow' — not by default, but at configuration time, from JavaScript too. (review, tested 2026-09-22 at d2f77a1)
- [Jevmind Review: The Reasons Report, They Don't Bind](https://mrjev.com/projects/dealerdefi-jevmind/): A decision gate, a hash-chained ledger and nine skills. Point it at a model and a confident wrong answer allows the command its own code just flagged. (review, tested 2026-09-22 at 308e320)
- [Canny Review: Only Facts Can Block](https://mrjev.com/projects/qkal-canny/): A warden that refuses an agent's 'done' when the ledger holds no passing check. The model can relax that refusal, and it can never cause one. (review, tested 2026-09-22 at a266600)
- [Abide Review: Your AGENTS.md Rules, Enforced](https://mrjev.com/projects/coldteadotai-abide/): Hooks into Claude Code, Codex and OpenCode and checks every edit against your own rules. It found what it should, but one privacy claim doesn't hold. (review, tested 2026-09-20 at ec3352e)
- [jevscan-evm Review: 359 Questions Per File](https://mrjev.com/projects/devtooligan-jevscan-evm/): A bug heat map for Solidity, from an author who says they did not read the code. Point it at a scope list and it sends whatever is on it. (review, tested 2026-09-20 at current)
- [jev-guard Review: Auto Mode for Any Agent](https://mrjev.com/projects/leepokai-jev-guard/): Risk-scores every tool call across eight agents and catches injections in results. Sub-agent delegation is exempt from both checks. (review, tested 2026-09-20 at current)
- [is-malicious Review: It Reads Your .env by Design](https://mrjev.com/projects/luantak-is-malicious/): Scans a codebase for hidden or data-stealing behaviour before you run it. `.env` is on the list of files it sends, and an outage passes the build. (review, tested 2026-09-20 at 240653e)
- [tripwire Review: Judging Every LLM Reply](https://mrjev.com/projects/noelzappy-tripwire/): Seven safety checks about a model's reply in one request, cheap enough to run on all traffic instead of sampling. Small, tested, and well posed. (review, tested 2026-09-20 at b727e6f)
- [jev-gates Review: Seven Gates That Only Escalate](https://mrjev.com/projects/rashedint32-jev-gates/): Hooks that ask before an edit, before a commit, and before Claude stops. It never approves — it only raises a hand — and every failure opens all the gates. (review, tested 2026-09-20 at bb89182)
- [jev-belay Review: Catching an Unverified 'Done'](https://mrjev.com/projects/valentynkit-jev-belay/): A Claude Code stop hook that only spends a call when the evidence looks wrong: files changed, nothing checked, and the agent says it's finished. (review, tested 2026-09-20 at 98f39e0)
- [Pi Jev Guard Review: The One That Fails Closed](https://mrjev.com/projects/zszz3-pi-jev-guide/): A Pi plugin that judges each tool call before it runs and redacts output after. Its local credential gate skips commands that also upload. (review, tested 2026-09-20 at 0.2.1)
- [pi-warden Review: Guardrails for the Pi Agent](https://mrjev.com/projects/devmortimer-pi-warden/): An ambitious, thoroughly documented guardrail layer for the Pi coding agent. Its offline safety net works well; our redaction fix is merged, not yet on npm. (review, tested 2026-09-19 at d015953)

## Reviews: Evaluation & Benchmarks

- [Jev vs ML Note: A Benchmark You Can Audit](https://mrjev.com/projects/quicqdev-jev-vs-ml/): Jev against eleven classical classifiers on eight datasets, with hashes and stated caveats. Checked, not fully reviewed: the tables match the saved run. (note, tested 2026-10-10 at 6c261e8)
- [jeval Note: What Your Confidence Is Really Worth](https://mrjev.com/projects/rlaope-jeval/): Calibration reports and cost-based thresholds from labelled decision logs, offline. Checked, not fully reviewed: careful maths, on a scale Jev does not use. (note, tested 2026-10-10 at 61bdccc)
- [Janus Review: Measure First, Then Decide to Route](https://mrjev.com/projects/firassx914-janus/): Sweeps every confidence threshold on your own labelled data and is willing to conclude DO NOT ROUTE. Since 0.3.2 its budget flag fails closed. (review, tested 2026-10-06 at 1144f3d)
- [JevBench Review: The Ruler Others Cite](https://mrjev.com/projects/fstandhartinger-jevbench/): Every ranked score recomputes from its published axes, and the datasets match their hashes. One constant needed to check the top axis is not written down. (review, tested 2026-09-27 at d06ee95)
- [jevals Review: The Extra That Weakens Redaction](https://mrjev.com/projects/openlayer-ai-jevals/): Agent evals as typed questions in one request instead of an LLM judge. Installing its PII extra stops phone numbers and SSNs being redacted. (review, tested 2026-09-24 at 0457836)
- [jev-calibrate Review: Good Numbers, Refused](https://mrjev.com/projects/smkrv-jev-calibrate/): Checks a Jev question against your own labels and grades it. We gave it a perfect score on three examples and it declined to bless the question. (review, tested 2026-09-22 at 28bc620)
- [Jev Capability Atlas Review: A Headline Withdrawn](https://mrjev.com/projects/zaious-jev-capability-atlas/): Six suites that ship every raw receipt. A typo we reported turned out to have produced its flagship result, and the project withdrew it. (review, tested 2026-09-21 at b85bae4)
- [jev-benchmarks Review: A Pre-Registered Comparison](https://mrjev.com/projects/abdelstark-jev-benchmarks/): Compares Jev with an open zero-shot classifier under a written protocol, with hashes and resumable runs. Auditable rather than reproducible. (review, tested 2026-09-20 at 0d610cc)
- [inference-benchmarks Review: Two Lanes, One Bench](https://mrjev.com/projects/iammrduncan-inference-benchmarks/): Runs the same seven scenarios through a frontier model's structured output and through Jev, and shows its working. Its own README has gone stale. (review, tested 2026-09-20 at cf348cd)
- [jev-align Review: Label the Rows It Isn't Sure Of](https://mrjev.com/projects/sutro-sh-jev-align/): Finds the rows Jev is least certain about, has you label them, and lets GEPA rewrite the question. It never measures the calibration in its name. (review, tested 2026-09-20 at 49753df)
- [jevcal Review: Pick Jev Confidence Thresholds](https://mrjev.com/projects/abhixhek-jevcal/): The tool to use before you automate on Jev's confidence: measures your own data, picks thresholds honestly, and fails CI on drift. Young, but well built. (review, tested 2026-09-19 at ae8f314)

## Reviews: Model Routing

- [stuntd Note: A Proxy That Learns Your Decisions](https://mrjev.com/projects/bladedevoff-stuntd/): Records your app's typed decisions and trains a local Laya head to answer them. Checked, not fully reviewed: honest about who answered; fresh installs fail. (note, tested 2026-10-11 at 4d90149)
- [Astra-Ares: Reasoning Effort, Chosen Mid-Task](https://mrjev.com/projects/miuuyy-astra-ares/): Lets a decision model raise or lower GPT-6's reasoning effort while a Codex task runs. Checked, not fully reviewed: it needs a patched Codex and a Mac. (note, tested 2026-09-27 at b201144)
- [Helm Review: It Cannot Route To What You Lack](https://mrjev.com/projects/jimuelle07-helm/): An orchestrator that picks between the coding agents already on your machine. The answer space is built from what it found, and we tried hard to break that. (review, tested 2026-09-22 at 2cdedba)
- [pi-jev-router Review: A Knee, Not a Weight File](https://mrjev.com/projects/philippdubach-pi-jev-router/): Five typed questions about a task, then a Pareto frontier over the OpenRouter catalogue picks the model. Shadow mode is the default. (review, tested 2026-09-21 at b9650e6)
- [jev-gateway Review: Jev Picks, the LLM Writes](https://mrjev.com/projects/vinilana-jev-gateway/): A local proxy that lets Jev choose the tool and closed-set arguments while the model writes the prose. The three edge cases we found are fixed. (review, tested 2026-09-20 at 9463952)
- [Safer with Jev Review: A Showcase You Can Poke](https://mrjev.com/projects/andrelandgraf-safer-with-jev/): A public demo that judges a request body and, if you ask, forwards the same bytes onward. Well tested, and the whole point is that you send it things. (review, tested 2026-09-20 at b4fb82b)
- [JevRouter Review: One Question, Every Capability](https://mrjev.com/projects/billionsbobby-jevrouter/): Routes a request across models, subagents, skills, MCP tools and CLIs with one Jev call. Its quickstart shape silently drops a candidate's risk level. (review, tested 2026-09-20 at 3c558a7)
- [grok-bot-jev Review: A Gate a Sentence Can Open](https://mrjev.com/projects/bodila51-grok-bot-jev/): Routes Grok Bot tasks through one Jev call and recommends an action. Its approval gate is disabled by a phrase a customer might write. (review, tested 2026-09-20 at 1583e09)
- [tiershift Review: Routing With a Reproducible Bench](https://mrjev.com/projects/iamvatsalpatel-tiershift/): Eleven signals in one request decide which model tier a request deserves — and its published benchmark recomputes exactly from the data in the repo. (review, tested 2026-09-20 at 16a0826)
- [jev-router Review: Per-Turn Routing for Claude Code](https://mrjev.com/projects/gargpratyush-jev-router/): A well-built local proxy that lets Jev pick the Claude or Codex model for each turn. Safe defaults, but read the privacy and compatibility notes first. (review, tested 2026-09-19 at 0d39e5b)

## Reviews: Community SDKs

- [laya-php Note: Typed Decisions as PHP Enums](https://mrjev.com/projects/marcreichel-laya-php/): A clean, well-tested PHP and Laravel client for Laya. Checked, not fully reviewed: since 1.2.1 it works with any server that follows TypeSafe's spec. (note, tested 2026-10-07 at 9a9c8ea)
- [PHP SDK for Jev Note: Attributes In, Answers Out](https://mrjev.com/projects/sanmai-typesafe-ai-php/): A small, tested PHP client for Jev's three answer types. Checked, not fully reviewed: reads the published shape, retries 529, sends objects' private fields. (note, tested 2026-10-06 at d71bb9e)
- [swift-typesafe Note: Typed Questions as Swift Types](https://mrjev.com/projects/ainame-swift-typesafe/): A Swift port of TypeSafe's Python SDK with macros for typed question sets. Checked, not fully reviewed: it runs on Linux, but option order is lost. (note, tested 2026-10-04 at 68f954b)
- [Spring AI TypeSafe Note: Jev for Spring Apps](https://mrjev.com/projects/spring-ai-community-spring-ai-typesafe/): A Java SDK, Spring Boot starter and Spring AI advisors for Jev. Checked, not fully reviewed: an offline suite of 286 tests, and how it fails under load. (note, tested 2026-10-04 at 2c30a58)
- [typesafe-sdk-java Review: Zero-Dependency Java](https://mrjev.com/projects/kgonia-typesafe-sdk-java/): A Java 21+ client with no dependencies at all, a module-path test, and 114 passing tests. Small, sober, and it builds first try. (review, tested 2026-09-20 at 459c285)
- [Swift TypeSafe SDK Review: Doesn't Build on Linux](https://mrjev.com/projects/krzyzanowskim-typesafe/): A small, well-shaped Swift client for Jev. It has no CI, and it will not compile on the server platform its own README recommends. (review, tested 2026-09-20 at 4195213)
- [kunobi-decision Review: The Rust Client, Well Tested](https://mrjev.com/projects/kunobi-ninja-kunobi-decision/): 94 tests, a fuzz target, a licence audit config and blocking plus async APIs — in a crate with no stars at all. The most thorough small SDK here. (review, tested 2026-09-20 at e151485)
- [typesafe_sdk Review: The Elixir Client, Renamed](https://mrjev.com/projects/nshkrdotcom-typesafe-sdk/): An Elixir client that has already been superseded by its own successor. Read this one only to understand where system_one_sdk came from. (review, tested 2026-09-20 at 31cd200)
- [.NET SDK Review: A Careful Community Jev Client](https://mrjev.com/projects/saibimajdi-typesafeai-dotnet-sdk/): A community .NET client for Jev with real depth: typed answers, mapped errors, retries and forward compatibility. Everything we threw at it behaved. (review, tested 2026-09-20 at 6be669e)
- [typesafe-go Review: A Go Client, Quietly Complete](https://mrjev.com/projects/shubham510-typesafe-go/): Two commits, no stars, and all three question types with typed errors and retries. Needs Go 1.27, which is newer than most people have. (review, tested 2026-09-20 at 47273ef)

## Reviews: Data & Observability

- [Jeview Review: A Gateway That Keeps The Key](https://mrjev.com/projects/andududu-jeview/): A local gateway that draws every Jev call your code makes. We sent our own bearer token through it, and it never reached the other side. (review, tested 2026-09-22 at 495a4e4)
- [ultralightspeed Review: All Of It, Or It Raises](https://mrjev.com/projects/collapseindex-jev-ultralightspeed/): Packs many items into one Jev request for bulk work. If any item in a pack comes back unanswered, it raises and names the item rather than guessing. (review, tested 2026-09-22 at f579487)
- [DocJev Review: One Request for Fifteen Pages](https://mrjev.com/projects/jerryjliu-docjev/): Classify and split documents with Jev. A 15-page packet and its 29 questions go out in a single call, and the benchmark recomputes from its own records. (review, tested 2026-09-22 at 7e6b48d)
- [jevframe Review: Only The Columns You Name](https://mrjev.com/projects/ktaletsk-jevframe/): A .jev accessor for pandas and Polars. We put canaries in the columns we did not name, and none of them left the machine. (review, tested 2026-09-22 at 16bd3ea)
- [duckdb-jev Review: Typed Decisions as SQL Types](https://mrjev.com/projects/colliber-duckdb-jev/): A DuckDB extension where a Choice becomes an ENUM column. Well built — but on a NULL or failed row, jev_ask hands back an answer that isn't there. (review, tested 2026-09-20 at fad67eb)
- [pg_typesafe Review: Jev From Inside Postgres](https://mrjev.com/projects/giuliosmall-pg-typesafe/): A C extension that calls Jev straight from SQL, with careful key handling. One setting lets any granted role forge every answer. (review, tested 2026-09-20 at 4b5bfc1)
- [jevQL Review: Semantic SQL Without an Extension](https://mrjev.com/projects/kylemclaren-jevql/): Rewrites jev() calls out of your SQL, judges the rows outside the database, and puts the answers back. Vanilla Postgres, no extension to install. (review, tested 2026-09-20 at 274532a)
- [tax-doc-classifier Review: 261 IRS Forms, No Model](https://mrjev.com/projects/kyotofin-tax-doc-classifier/): A production-shaped page classifier that names IRS forms with one Jev request. It worked on every page we tried, but read `gated`, not `form`. (review, tested 2026-09-20 at 6afcf70)
- [doc-router Review: Don't OCR a Page That Has Text](https://mrjev.com/projects/misbahsy-doc-router/): Rust that reads a PDF page by page and sends only the pages that need OCR. On its own trap corpus, the Jev judge caught what the heuristic missed. (review, tested 2026-09-20 at 8977d2b)
- [jevlogs Review: Scoring Logs Before the LLM](https://mrjev.com/projects/reachjalil-jevlogs/): Scores each log line for diagnostic value, priority and routing so only the interesting ones reach an expensive model. Careful about secrets, too. (review, tested 2026-09-20 at b1ff600)

## Reviews: Finance

- [BTC Decision Lab Review: Live Mode Is Refused](https://mrjev.com/projects/frankda-jev-poly-crypto-demo/): A Polymarket research tool that will not place a real order: any TRADING_MODE other than paper throws at startup. We tried it. (review, tested 2026-09-22 at a017b84)
- [jev_stock Review: A Backtest That Doesn't Leak](https://mrjev.com/projects/sosopop-jev-stock/): Forecasts Hong Kong share direction from a past-only state. We planted future prices in the series and none of them reached the request. (review, tested 2026-09-21 at 81fa919)
- [jev-trade Review: Five Bots, No Risk Controls](https://mrjev.com/projects/aowang-ai-jev-trade/): Asks Jev for direction, intent and leverage every two seconds across five assets, then trades it. There is no stop, no cap, and no kill switch. (review, tested 2026-09-20 at 2d5f002)
- [Jev X Sentiment Review: Confidence Theatre](https://mrjev.com/projects/brainstormity-jev-x-sentiment-analysis/): A crypto terminal that reads a thousand posts and prints a trade ticket. It cannot touch your money — but the price on that ticket may be invented. (review, tested 2026-09-20 at 5c932f9)
- [Prism Review: Jev on the Sidelines of a Trading Bot](https://mrjev.com/projects/irfndi-prism-liquidity-agent/): A Solana liquidity agent that keeps Jev advisory and paper-only. Serious engineering, and one remote endpoint it trusts more than it should. (review, tested 2026-09-20 at 22c67bd)
- [jev-trader Review: A Jev Market-Making Bot on Monad](https://mrjev.com/projects/jarrodwatts-jev-trader/): A clean, well-documented demo of Jev making a trading call every Monad block. Safe to try in dry-run mode; live mode has no loss limits. Not financial advice. (review, tested 2026-09-19 at b587759)

## Reviews: Search & Knowledge Graphs

- [Jev Search Review: Fifteen Questions, One Request](https://mrjev.com/projects/superagents-lab-jev-search/): Plain-language web search where Jev picks the sources, the time range and the query in one request, then scores every result. No generated answers. (review, tested 2026-09-22 at 67027d0)
- [blink Review: Walking a Tree by Name](https://mrjev.com/projects/ellipsis-dev-blink/): Sends the names in a directory and asks which one probably holds the answer. Clever, tiny, and it will happily write 83 MB before calling anything. (review, tested 2026-09-20 at a621ede)
- [neo4jev Review: Walking a Knowledge Graph With Jev](https://mrjev.com/projects/jexp-neo4jev/): A Neo4j demo that beam-searches a graph by letting Jev score each edge. It found what we asked for, and it is honest about what it samples. (review, tested 2026-09-20 at d157bbe)
- [jev.nvim Review: Finding the Function That Does It](https://mrjev.com/projects/valentynkit-jev-nvim/): Splits a buffer into functions with treesitter, asks one question each, and ranks them into the quickfix list. Its accuracy table is deliberately blank. (review, tested 2026-09-20 at a1577b0)

## Reviews: Robotics & Embodied

- [jev-libero Review: Previews First, Then a Pick](https://mrjev.com/projects/dimweaker-jev-libero/): Simulates every candidate move locally, then asks Jev to choose one. Its recorded run reproduced on our machine, decision for decision. (review, tested 2026-09-21 at 9b8098a)
- [EmbodiedJev Review: A Robot Lab in a Browser](https://mrjev.com/projects/fbddcz-embodied-jev/): Three MuJoCo tasks, a rule baseline, and a confidence gate that stopped the arm mid-task and waited for us. Careful with keys, Chinese-only docs. (review, tested 2026-09-21 at 59a00c6)
- [Jev Reflex Lab: Drones That Ask for Advice](https://mrjev.com/projects/khordoo-jev-reflex-autonomy-lab/): A fleet simulation where typed reflexes fly and a slower planner may advise but never steer. Good guards; no licence, and its own tests fail. (review, tested 2026-09-21 at 22fd326)
- [RoboJEV Review: A Robot Arm That Only Picks](https://mrjev.com/projects/lykycy123-robojev/): Jev drives a Franka arm in MuJoCo from structured state, never pixels. Careful about what it claims, and the whole loop runs without a key. (review, tested 2026-09-21 at aa82be0)

## Optional

- [Full text for language models](https://mrjev.com/llms-full.txt): every review, guide and changelog entry as Markdown
- [About MrJev](https://mrjev.com/about/): MrJev is an independent, unofficial resource about Jev. Who runs it, how we source information, and how projects get listed. (updated 2026-09-18)
- [Privacy Policy](https://mrjev.com/privacy/): What data MrJev collects (cookieless analytics and newsletter emails only), why we collect it, and how to have it removed. (updated 2026-09-18)
- [Changelog RSS](https://mrjev.com/rss.xml)
- [Sitemap](https://mrjev.com/sitemap-index.xml)
- [awesome-jev on GitHub](https://github.com/MrJev/awesome-jev): the curated list the directory is built from