Your agent stops grepping. It starts answering.
Source-cited repo context, injected the moment you hit Enter.
Website · Docs · Benchmarks
Same question, same repo, same Codex. 36 s with Pixel. 1 min 58 s without. One recorded run, sped up uniformly — source and limits.
curl -fsSL https://github.com/Pixel-CLI/pixel/releases/latest/download/install.sh | sh
pixel setup # pick your agentsRestart your agent, approve hooks if asked, ask your next code question. That's it: Pixel indexes the repository on its first use. Want it warm up front? pixel prepare-repo .
Read the installer · release notes · build from main for unreleased integrations.
Your agent burns its first turns on discovery: plan a grep → run it → read a 2,000-line file → grep again → repeat. You pay for every loop. In tokens. In wall time. In context it never uses.
Pixel hooks your agent's prompt-submit event. Before the model sees your question, Pixel queries a local code index and attaches a compact [PIXEL:BRIEF]: definitions, callers, code, tests and git history, cited file:line.
The agent starts with the evidence already in context. It keeps all its tools — it just doesn't have to go hunting.
pixel brief "Where is request validation implemented?" . # see what your agent gets
pixel doctor . # check install and index health
export PIXEL_BRIEF=0 # turn briefs offNeed more than the brief? pixel setup drops a pixel skill into each agent's skills/ folder. The agent asks the index directly — one call, then a targeted read:
| Agent asks | Pixel answers with |
|---|---|
| where is X defined? | pixel find-symbol 'X' |
| who calls X? | pixel who-calls '<uid>' |
| what breaks if I change X? | pixel impact '<uid>' --depth 1 |
| outline of this 2,000-line file? | pixel list-signatures <file> — median −94.5% read volume |
| just this one function | pixel pack-context '<uid>' --budget 1500 |
Both animations illustrate a workflow; they are not timed agent comparisons. Agent trials and their limits.
| without Pixel | with Pixel | |
|---|---|---|
| Same Codex question on Requests (one recorded run) | 1 min 58 s | 36 s |
Requests models.py: full read vs. outline |
10,365 tokens* | 641 tokens* (−94%) |
| 8 large open-source files: full read vs. outline | 100% | median 5.5% (−94.5%) |
*Estimated tokens (bytes ÷ 4), Pixel 0.8.0 (the Codex run, Pixel 0.5.x), no billed cost. Every number has a protocol, and we publish the losses too: benchmarks.
Keeping the agent's context bill down is an idea whose time has come — teams at AWS and Spotify, and projects like xerj, ship their own takes on it:
| What it does | Mechanism | Cost model | |
|---|---|---|---|
| Pixel | Attaches cited repo context (file:line brief: definitions, callers, tests, git history) before the model sees your prompt |
Local code graph + prompt-submit hook | Deterministic index lookup — no model call, no latency, free |
| AWS AgentCore Memory sample | Wraps every query(): fetches short-term history and long-term summaries before the call, saves turns after |
Claude Agent SDK + managed AgentCore Memory | Managed AWS service; LLM-extracted cross-session memory, not code structure |
Spotify shunt |
Hooks block the agent from reading large files directly; a skill delegates the read to a cheaper internal LLM that summarizes (82–94% token savings claimed) | Claude Code PreToolUse hooks → AiKA modes via Portal CLI |
An extra billed model call per delegated read or codegen |
| xerj | Indexes any folder (code through tree-sitter, plus docs, logs and PDFs) into a local Elasticsearch-compatible engine that the agent searches instead of reading files (2.7× fewer output tokens than grep claimed on reference coding) | Local server + xerj search / xerj def CLI, or the MCP tool and skill xerj init installs |
Free and local, but a server to keep running and a search call per lookup |
We put a xerj brief (its top passages for the prompt, through the same hook and at the same size) in place of Pixel's on 13 arena prompts in Claude Code: 26 and 25 tool calls with Pixel's brief in two runs, 37 with xerj's. Installed with xerj init (MCP tools and a skill), the agent never called xerj (58 calls, none to xerj). Method, per-prompt results and indexing times.
Same goal, different lever: AWS and Spotify delegate work to another model or service, and xerj gives the agent a search engine it still has to call. Pixel makes the context a deterministic lookup in a local index, attached before the prompt. No extra tokens, no extra round trip, no cloud dependency — the same brief every time, cited to the line.
Claude Code · Codex · pi · Devin · OpenCode · Gemini CLI · Antigravity
One pixel setup wires each host's hooks or extension.
- 🦀 Single Rust binary · macOS (Apple Silicon) and Linux
- 🏠 Indexing and lookup run on your machine — no account, no API key, no telemetry
- 📜 MIT licensed · see NOTICE for attribution
Typed decisions from a local model — routing, gating, triage — one probability per label:
pixel classify "Review this Rust PR, find correctness bugs, and propose a safe patch." \
--engine ollaya \
--context "Choose the cheapest model that can reliably handle the request." \
--label 'claude-haiku-4.5_(fast)' \
--label 'claude-opus-5.5_(strong)' \
--label 'claude-fable-5.1_(reasoning)' \
--criterion 'claude-haiku-4.5_(fast)=Simple rewriting, extraction, or classification; no deep reasoning.' \
--criterion 'claude-opus-5.5_(strong)=Complex coding, multi-file review, or tool use; accuracy matters.' \
--criterion 'claude-fable-5.1_(reasoning)=Multi-step analysis, difficult debugging, or high uncertainty.'
claude-fable-5.1_(reasoning): 0.192
claude-haiku-4.5_(fast): 0.098
claude-opus-5.5_(strong): 0.710
predicted: claude-opus-5.5_(strong)Runs locally with Ollaya, or remote through OpenRouter, Ollama, DeepSeek, OpenCode Go, OpenAI or any OpenAI-compatible endpoint. Model bake-off →
Docs · Manual setup · Benchmarks · GitHub Actions · Architecture · Contributing · Security · Report a bug
