DA`/50 · open-source · local-first

Run commodified intelligence
at industrial volume.

Intelligence has become a commodity: cheap, abundant, interchangeable. That makes it a pipeline problem, not an app problem. Fifty-cent models at ten-billion-token scale need plumbing that is stateless, composable, and controllable. DA`/50 is that plumbing: a unix-grade call layer, memory that travels between context windows, agents with real control flow, and spend that cannot overshoot.

Start with llcat Browse the cabinet
100B tokens/day is the operating regime these tools are built for
1% saved is a billion tokens a day. Context reuse is the largest lever
0 servers to rent, nothing sent to us. Local-first by design
1 command to reach any model, any protocol, any provider

The front door

llcat: /usr/bin/cat for LLMs

An introduction. One stateless, transparent command that talks to any OpenAI-compatible endpoint: OpenRouter, Ollama, llama.cpp, sglang. Nothing is installed beyond the tool itself.

  • Universal. Keys, models, servers and conversations are explicit arguments. Rotate keys from a file (-k @credentials.txt:0), batch across models, run N×M×P jobs in parallel.
  • Stateless. No config files, no caches, no leaky state between runs. --curlify shows you the raw request. Nothing is hidden because nothing is the Emperor's clothes.
  • Composable. It is the base layer everything else in this cabinet runs on or pairs with. Master the call and the rest is within reach.
Install llcat Docs
llcat 
$ uvx llcat openrouter.ai/api -m
  OpenRouter model list:
    openai/gpt-5.5
    anthropic/claude-sonnet-4-6
    deepseek/deepseek-v4
    qwen/qwen3-235b-a22b
    meta-llama/llama-3.3-70b

$ for i in {0..40}; do llcat -k @credentials.json:.[1].password -m @models.txt:$i \
    "evaluate this prompt against 40 models, in parallel"; done
  → 40 responses, 3 servers, one line

The other door

ursh: enterprise compliance for what you run

Write your enterprise's compliance document once. SOC-2, HIPAA, favorable-license posture, whatever you must stand behind. Then audit any piece of software: a script, a dependency chain, a package from npm, PyPI, gem, or crates. Find out immediately whether it is compliant with your enterprise.

  • The compliance document is the source of truth. Every audit produces a machine-readable permission manifest; ursh answers one question against your policy: compliant, or not.
  • Transitive dependency analysis. Audit each dependency and use set operations: union, intersect, difference to reveal the real permission surface of the whole chain.
  • Policies you can enforce. Generate policy files from your compliance document and gate every future run on them.
Install ursh Docs
ursh 
$ ursh policy --from soc2.yaml
  compliance doc loaded: soc-2 · 12 rules

$ ursh audit gh:org/ci/deploy.sh
  ✓ passes policy

$ ursh audit gh:vendor/telemetry.sh
  ✗ NOT COMPLIANT
    network.put https://telemetry.example.com/events (line 41)
    not permitted by soc-2 policy

How it stacks

One system, not a pile of tools.

Each stage is a small, composable, MIT tool. Together they form the operating layer: feed the machine, control the agent, remember the context, count the cost, prove the trust.

  1. 01

    Talk

    Reach any model, any protocol, any provider.

    llcat · llsed · MAS

  2. 02

    Remember

    Stop re-burning tokens. Move memory between context windows.

    ctools · gabngo

  3. 03

    Control

    Keep agents on task. The model is the ALU, not the driver.

    structured-skills · haberdash · maxac

  4. 04

    Spend

    Caps that are enforced by the provider, not aspirational.

    capit

  5. 05

    Trust

    Enterprise compliance documents: audit any script or dependency against your policy.

    ursh

  6. 06

    Serve

    Meet humans where they work: the terminal.

    streamdown · sidechat · zummoner · ESChatch

  7. 07

    Feed

    Crawl, dedup, and preprocess corpora before a single expensive token.

    wgetjs · undecorate · ono · dailyllm.news

  8. 08

    Scale

    Many agents, one codebase: merge-safe filesystems, no collisions.

    agentfs · StackedDiffFS

The cabinet

Every drawer.

Talk to models

llcat

cat for LLMs. Stateless, composable, scriptable.

py

llsed

Proxy that translates between LLM API formats on the fly.

go

MAS

Name any model in a URI fragment. 10 reference implementations.

spec

Memory & context

ctools

mtools for context windows: cdir, ccopy, cdu, cgrep, cconnect.

py

gabngo

Context window as a filesystem: push and pull concepts.

py

Agents & control flow

structured-skills

A tiny VM where the LLM is the ALU. Real control flow, DAP debugging.

py

haberdash

Runtime for agentic instructions. Cloning a repo and running a macro are one op.

py

maxac

Single-shot agent: define success, plan, execute, verify, learn.

py

agent-cli-helper

Makes LLMs, even small ones, use interactive terminal apps.

sh

olm

A text file of one-liners becomes a full MCP server.

py

Spend & trust

capit

Budget per-agent, per-provider. Keys that literally cannot overspend.

py

ursh

Enterprise compliance: audit software against your compliance document.

go

Parallel filesystems

agentfs

Merge-safe overlay filesystem, every agent works in its own layer.

py

StackedDiffFS

Layered FUSE with pre/post hooks. Redact, log, filter, merge.

py

The human front

streamdown

Realtime streaming markdown renderer for the terminal.

py

sidechat

tmux AI assistant at the point of failure.

sh

zummoner

Summon the LLM at your shell. English in, real command out.

sh

ESChatch

Inject LLM keystrokes into any pty: vim, ssh, python.

py

llwhat

Ask your man pages questions. Local, semantic, private.

py

Feed the machine

wgetjs

wget for JS sites: headless Chrome, MIME-scoped post-processing.

py

undecorate

Strip shared boilerplate lines out of markdown corpora.

py

ono

Universal AI preprocessor. Write once, deploy in any language.

py

dailyllm.news

Cron pipeline: scrape, dedup with a local LLM, ship an RSS feed.

sh

Standards & teaching

concept

Hack-day workshop: local LLMs with llama.cpp, seven modules.

docs

Incubator

Early-stage HCI experiments: shell snoopers, Xorg interceptors.

misc