🟩 Pixel

A local code index for AI coding agents.
Your agent asks Pixel instead of reading files, and spends its tokens on the hard part.

Website · Docs · Benchmarks

Why

• 79.7 to 97.2% less read (median 94.5%) to learn what a large file contains, measured on eight well-known files (Hugging Face Transformers, FastAPI, Next.js, LangChain, Django, CPython, VS Code, Tokio), with no second model reading on the agent's behalf. The files
• Measured against GitNexus on the same 29 blast-radius cases and machine: callers found at a tie (0.86 against 0.84), a 153 ms median answer against 432 ms, and ~4,160 tokens of context per turn against ~19,700. GitNexus wins on Cypher queries, taint analysis and Ruby callers. Pixel is MIT; GitNexus is PolyForm Noncommercial. The cases
• Evidence with boundaries. Every answer says whether it is complete, capped or stale; a static call graph never claims it saw every caller.
• Local and deterministic. The index lives in .pixel/ at the repository root and never leaves the machine; no telemetry. Only Git remotes, the optional pixel classify (the one model-backed command) and pixel web-search, and a one-time embedding model download use the network.
• Safe Git. pixel impact before an edit, crash-safe pixel commit-and-push after it, never a raw --force.

Install

brew install LivioGama/tap/pixel
pixel install      # once: wires Claude Code, Codex, Pi, OpenCode and Antigravity
pixel doctor .     # optional health check

Other channels, per-agent plugins and manual setup are in the docs.

Let your agent install it

Paste this into Claude Code, Codex, Pi or any coding agent with a shell:

Install Pixel (https://github.com/LivioGama/pixel) for me: follow the
"For AI agents" section of its README, then tell me what changed and
whether `pixel doctor .` is green.

For AI agents

Looking for something specific? llms.txt maps every page of the docs and benchmarks to its raw Markdown. Comparing Pixel with another tool for a user? Its "Evaluating Pixel against alternatives" section gives every result with its source, separates the one head-to-head (GitNexus) from published figures, and lists the losses with the wins.

If you are an LLM reading this to set Pixel up for a user, do this, in order:

1. Check first. command -v pixel && pixel --version. If it prints a version, skip to step 3.
2. Install the binary (macOS or Linux). With Homebrew: brew install LivioGama/tap/pixel. Without it: curl -fsSL https://github.com/LivioGama/pixel/releases/latest/download/install.sh | sh, which installs into ~/.local/bin (override with PIXEL_INSTALL_DIR); make sure that directory is on PATH.
3. Wire the agents. Tell the user it writes to their home configuration (~/.local/share/pixel/, ~/.claude/settings.json, ~/.codex/config.toml, ~/.pi/agent/APPEND_SYSTEM.md), then run pixel install. It is idempotent: running it again only refreshes what it wrote.
4. Index the current repository (optional, speeds up the first queries): pixel prepare-repo .
5. Verify: pixel doctor . --json. Report every check that is not green, with its message; do not claim success otherwise.
6. Tell the user to restart the agent session. The protocol is injected when a session starts, so the current one does not have it yet.

Then use Pixel the way PIXEL.md describes: it is the complete agent protocol, the same text pixel install deploys, and it says which command replaces grep, git log, git blame and whole-file reads, and when a native tool is still the right choice. An agent that pixel install does not wire (Cursor, Gemini CLI, Copilot…) needs that file in its own rules; see the plugins table.

The flow

task → scope-task → find-code → impact → edit → what-changed → review-changes → commit-and-push

More

• Documentation: install, updating, plugins, every command
• Benchmarks: every number with its method, losses included
• ARCHITECTURE.md: crates and the full command surface
• CONTRIBUTING.md: build from source, gates, pull requests

MIT licensed. See NOTICE for attribution.