ctx — token-saving context engine

Local middleware between your AI coding tool (Claude Code, Cowork, Codex, any MCP host) and the model. It shapes every request to be minimal but intelligent: a graph-ranked repo map instead of whole files, an append-only decision log instead of replayed history, a stable cached prefix, and diff-only output. Saving tokens and keeping the model sharp are the same goal — see ARCHITECTURE.md.

Install

pipx install token-diet            # core (runs with graceful fallbacks)
pipx install 'token-diet[all]'     # full: MCP + proxy + tree-sitter + tiktoken + docs

Not on PyPI yet? Install straight from GitHub:

pipx install 'git+https://github.com/aryxnsdfs/token-diet'

From source:

pip install -e '.[all,dev]'

The core works with zero heavy deps via fallbacks (regex parser, char-based token estimate, heuristic distillation). Install extras to upgrade each piece; ctx doctor shows what's active.

Related MCP server: token-pilot

Quick start

cd your-project
ctx init        # build index, register with host, write command files
ctx doctor      # verify wiring + see which optional deps are active

In Claude Code, press / and pick a command. First run /init (or /ctx start) to warm the chat — injects the repo map, enables diff mode.

Commands

Two front-ends, one engine

• Mode A — MCP server. ctx serve exposes every command as an MCP prompt (you type /name) and tool (the model calls it). The native path for MCP-aware hosts.

• Mode B — local proxy. ctx proxy runs a localhost:8000 server you point a client at; it optimizes every request and parses /text for non-MCP tools.

Layout

ctx/
  cli.py        init · index · serve · doctor · proxy
  registry.py   single source of truth for commands
  server.py     FastMCP: prompts + tools + resources
  proxy.py      Mode B FastAPI proxy
  init/         per-host registration adapters
  engine/       index · repomap · assembler · budget · memory · docs · patch · router · telemetry
tests/          unit + eval harness (the guardrail)

The guardrail

tests/evals/ holds representative tasks with automatic checks. Token reduction only counts as saving intelligence if task success holds. Run pytest on every engine change.