Quor

Your AI coding assistant is burning tokens on noise. Quor cuts it before it ever reaches the model. An MCP-native, zero-heuristic context-compression server. Runs entirely on your machine. No LLM, no cloud, no network call — just a deterministic rule pipeline that strips the boilerplate out of everything your assistant reads, exposed as standard MCP tools any MCP-compatible client can call.

35.9% smaller, on average. Up to 89% smaller on the worst offenders.

Measured across a 153-case hand-curated benchmark suite, CI-gated on every single change — not a one-time demo number. Benchmark cases are realistic, hand-authored samples, not a random draw of real usage — run quor gain / quor dashboard on your own project for what Quor actually saves you, or see the numbers below for the full benchmark breakdown.

Install — 30 seconds

pip install quor
quor init --mcp

This writes ./.mcp.json, registering Quor as an MCP server for the current project, and prints the equivalent claude_desktop_config.json snippet for Claude Desktop or any other MCP client. Requires Python 3.11+.

To upgrade: pip install --upgrade quor — nothing else to re-run; there's no launcher script to go stale.

If you have a pre-0.6 install with quor init --claude's old hook files still on disk, quor init cleans them up automatically the next time you run it (or run quor uninstall-hooks directly).

Why this matters

Every command your AI assistant runs — git status, a pytest run, a file read — pours straight into its context window. Most of it is boilerplate: passing tests it doesn't need to see, unchanged diff context, repeated warnings, PDF page furniture, dependency-install spam. That's tokens you're paying for, latency you're waiting on, and context-window budget your assistant isn't spending on your actual code.

The obvious fix — have the AI summarize its own output — is the wrong one: it doubles latency, doubles cost, and can silently drop the one line that mattered. Quor takes the other path: a local, rule-based pipeline that runs in milliseconds, makes the same keep/drop decision every time given the same input, and never touches the network.

assistant reads a command's output → calls Quor's compress_context MCP tool on it → rules mark each line KEEP / COMPRESS / PROTECT → noise drops → the assistant gets fewer tokens back

Same output, same meaning — only what reaches the context window changes. Failures, diffs, and tracebacks are never touched, and every compressed output links back to the full original — nothing is ever lost, just deferred until you ask for it.

MCP Tools

Quor runs as a standard MCP server (quor/mcp/server.py, stdio transport) exposing two tools:

The calling assistant decides when to invoke compress_context — on a large command output, a big file read, anything it judges worth compressing before it lands in context. This is the one real trade-off of moving to MCP: compression is opt-in per call, not a transparent interception of every command the way the pre-0.6 hook-based integration was. In exchange, Quor works with any MCP-compatible client instead of one hook implementation per assistant.

The Numbers

35.9% average token reduction across Quor's own 153-case, hand-curated benchmark suite (CI-gated — a regression here fails the build, not just a dashboard). That average includes plenty of already-terse output with nothing left to cut; on the cases that actually have noise to remove, it goes much further:

By ecosystem:

Short, already-dense output compresses little — that's correct behavior, not underperformance; Quor never trims a line just to move the number (see ANTI_GOALS.md). Full breakdown in docs/BENCHMARKS.md, or run quor gain / quor dashboard for your own project's real, live numbers — always shown with an honest ±20% uncertainty band, never a bare number dressed up as exact.

Why did my quor gain/quor dashboard percentage swing a lot between two runs? It's a running average across every command Quor has seen in the window, not a per-command score — one big compressible file read can dominate a small sample (86% after 2 commands), and the ratio settles as more commands, including plenty that are already small or have nothing Quor can filter (ps, grep, a one-line git diff), run afterward (7% after 45). Both numbers are correct the whole time. quor gain's "On the N% of commands a filter could apply to" line, and the Passthrough count both commands show, isolate the compression rate on just the content that was actually eligible — usually the steadier number to watch.

Commands

Full reference: quor --help.

Supported

Any MCP-compatible client can register and call Quor's tools — Claude Code, Claude Desktop, and every other client speaking the Model Context Protocol, with no per-assistant integration work required. quor init --mcp scaffolds registration for Claude Code (.mcp.json) and prints the equivalent claude_desktop_config.json snippet; any other MCP client follows the same mcpServers shape.

Commands the filter pipeline understands: git, pytest, mypy/ruff, pip/poetry, the full Node/TypeScript toolchain (npm, pnpm, yarn, ESLint, tsc, Jest, Vitest, Prettier, Next.js, Turbo), and a generic fallback for everything else. Source code: Python built in; JavaScript/TypeScript, Go, Rust, Java, C# via pip install "quor[<language>]". Documents: Markdown, TXT, DOCX, PDF via quor[documents]. Config: JSON/TOML/.env/.ini built in, YAML via quor[yaml].

Trust

An MCP server sees whatever text an assistant hands it for compression — it has to earn the right to be there.

• Local execution only — no network calls, no cloud, no telemetry, no API keys, ever
• Rule-based, not AI — pattern match, dedup, count, budget; zero ML in the filter path, zero hallucination risk
• Fail-open — a filter bug never blocks a command or hides output, it just returns the original untouched
• Secret-aware — warns (never silently strips) if a credential pattern survives compression
• Meaning-preserving by contract — a line Quor keeps is bit-for-bit identical to the original; nothing is rephrased or summarized (see ANTI_GOALS.md)
• App-control friendly — the MCP server runs via python -m quor.mcp.server directly, never an unsigned launcher .exe, so corporate AppLocker/Defender policies don't get in the way

Contributing

git clone https://github.com/priyanshup/Quor.git && cd Quor
pip install -e ".[dev]"
pytest tests/

CONTRIBUTING.md · SECURITY.md · CHANGELOG.md

License

Apache 2.0 — see LICENSE