OMNI

Stop paying Claude to read 10,000 lines of terminal noise. Over one developer's real week, OMNI cut 88% of build and test output and a quarter of everything the agent re-read, 15.7% across the whole mix. The other 97% of calls passed through untouched. Nothing is ever lost, and it never invents a result.

🇺🇸 English | 🇯🇵 日本語 | 🇨🇳 简体中文 | 🇸🇦 العربية | 🇮🇩 Bahasa Indonesia | 🇻🇳 Tiếng Việt | 🇰🇷 한국어

build and test 88% · file re-reads 25% · 15.7% across the mix · 21 ms per command · 2 of 7,095 calls grew the output, and we say so · every cut recoverable byte for byte · cross-session memory

brew install fajarhide/tap/omni && omni init

Distills command output on Claude Code, Codex CLI and Gemini CLI, where the host applies OMNI's rewrite. Everywhere else you get the MCP server, shared session state, and omni_run, which distils any command you route through it. Run omni doctor to see which tier each installed host is on.

What each host lets OMNI do

Tier	Hosts	What you get
Full	Claude Code, Codex CLI, Gemini CLI, Aider (pipe)	The host applies OMNI's rewrite, so the model reads distilled output from its own built-in tools.
Handoff-first	Cursor, Windsurf	The host cannot rewrite built-in tool output. omni_run distils anything you route through it, and omni init --cursor installs the rule that makes the agent reach for it.
MCP-only	Cline, Roo, OpenCode, VS Code, Zed, Copilot, Antigravity, Hermes, Pi	Memory, recall and session state. No shell distillation, and no claim of it.

omni doctor prints the tier for every installed host. Savings are only ever counted where the model actually received less.

Codex CLI needs one extra step. It runs only hooks it has been told to trust, and skips the rest without a word, so after omni init --codex start codex once and approve them under "Hooks need review". omni doctor fails until you do. See #359.

Your agent reads every line your terminal prints. Build logs, Docker logs, CI logs, progress bars, ANSI colors. Thousands of tokens to find one line. Claude isn't expensive. Your terminal is.

And it forgets all of it overnight. Restart Cursor, switch to Claude Code, and you re-explain the project from scratch.

OMNI fixes both, and stays out of the way everywhere else.

Without OMNI, one commit's Author / Date / body already fills the screen. With OMNI, every commit is kept, as one hash subject line, 94% smaller. Nothing is summarised away.

Without OMNI raw git log -15	With OMNI every commit kept, 94% smaller

Where it does nothing, on purpose. A command that fails is passed through verbatim, because a hidden error costs more than an uncompressed one. Structured output (JSON, YAML, CSV) is never touched, because the next step in your pipeline is going to parse it. That is what makes it safe to leave on for every command you run.

Two promises, and both are in the code rather than in this paragraph.

Nothing is ever lost. Every byte OMNI cuts is archived locally in the RewindStore, keyed by SHA-256. The agent gets a hash with the distilled output and can call omni_retrieve to pull the original back byte for byte, mid-conversation, without re-running your command.

It never makes something up. A distiller that recognises nothing in its input returns the raw input. That is a type, not a convention: distill returns Option<String> and the routing layer falls back to the original whenever it gets None. There is no code path that produces a green "no errors" line OMNI did not read.

Measured on the release binary by replaying 7,095 real command executions covering 2026-08-03 to 08-10 UTC, every one of them output that reached a model. The window is part of the figure: execution_traces is pruned after seven days, so a corpus is gone a week after it is measured.

• Where there is noise, the filters take almost all of it. Build and test output is 87.9%, and 92.3% once the session ledger is counted. Where there is no noise they take nothing, and a kubectl get pods table is 0%, because every row in it is a datum.
• The ledger reaches what filtering cannot. File re-reads are the largest class at 1.54 MB, the filters take 0.0% of them, and handing back lines the agent has already been shown takes 24.6%.
• 97.1% of calls saved nothing at all and handed the output straight back. Every byte of the saving comes from the other 2.9%.
• 2 calls of 7,095 came back larger, reported rather than rounded away (#398).
• 15.7% fewer bytes across the whole mix, of which the filters are 5.2% and the ledger is the rest. Counted in tokens the filters alone are 5.0%.
• 21 ms per command end to end, growing with your history rather than with the payload. On a 205 MB database it is 61 ms.

Full corpus, per-class breakdown, fixtures and latency tables: docs/BENCHMARKS.md. Reproduce them with cargo test --release --test bench_replay -- --ignored.

Every tool in this category publishes a percentage. Here are the five questions that decide whether it means anything, and our answers:

We publish the share of calls where we did nothing because it is the number that tells you what the rest are worth.

macOS / Linux:

brew install fajarhide/tap/omni
omni init      # interactive setup for Claude, Cursor, VS Code, Codex, Antigravity
omni doctor    # verify, or `omni doctor --fix`

Universal (macOS / Linux / WSL):

curl -fsSL omni.weekndlabs.com/install | bash

Windows (PowerShell):

irm omni.weekndlabs.com/install.ps1 | iex

Then run your commands normally. There is nothing to prefix and no proxy to wrap.

Does OMNI permanently delete my logs? No. Raw logs are stored locally in the SQLite RewindStore. The agent receives a hash and can retrieve the full log at any time.

Will this slow down my terminal? Measurably, yes, and the cost grows with your history rather than with the payload. A 496-byte git status takes about 21 ms against a fresh database and 61 ms against a 205 MB one. Budget for it. OMNI_PASSTHROUGH=1 skips the pipeline entirely.

Can I add my own filters? Yes, in TOML:

# ~/.omni/signals/custom.toml
[filters.my_tool]
match_command = "^internal-tool\\b"
strip_lines_matching = ["^DEBUG", "syncing..."]

How do I see my own savings? omni stats after a few days. omni stats --share prints a copy-pasteable summary of the same figures.

• How it works, and what it costs: pipeline, RewindStore, the Memory OS
• Benchmarks in full: corpus, per-command, fixtures, latency
• Contributing: make ci and you're in

brew install fajarhide/tap/omni && omni init

A passion project for the era of agentic AI. Contributions welcome. MIT License.

Built with care by Fajar Hidayat