MCP-native · Apache 2.0 · Self-hosted

code context
for AI dev tools

karst indexes your repo, packs it into small token-efficient bundles, and serves them over MCP to Claude, Cursor, and any AI dev tool. Local-first and self-hosted — with a local model, your code never leaves your machine.

Install free View on PyPI

343s → 2.3s

Re-index after first run. Incremental, content-addressed.

$0.019

Avg cost per query on Sonnet 4.6. Shown before each call.

60%

Token reduction with packs vs. raw file dumping.

How it works

Three steps. No vector DB to babysit, no embeddings drift, no surprise bills. See the interactive walkthrough →

01 / INDEX

Parse the repo

karst walks your tree, parses each file with tree-sitter, and writes a content-addressed index of symbols, calls, and imports. The first pass takes 343s on a 2M-line repo; subsequent runs are 2.3s because only changed blobs are re-parsed.

02 / PACK

Group into packs

Files that change together and call each other get grouped into packs — small, named bundles like auth or billing-stripe. karst packs suggest proposes them; you review, tag, and apply. Packs are the unit of context.

03 / SERVE

Serve over MCP

karst exposes packs as MCP resources. Claude Desktop, Cursor, and any MCP client can pull just the pack they need — 60% fewer tokens than dumping raw files, with the exact cost printed before each call.

Install in 60 seconds

Python 3.10+. Works on macOS, Linux, WSL.

# 1. install
pip install karst

# 2. index your repo (one-time, ~5 min on a large codebase)
karst index ./my-repo

# 3. suggest packs, review, apply
karst packs --storage ~/.karst/my-repo suggest ./my-repo --apply --retag

# 4. ask anything — cost is printed before the call
karst ask "How does auth work?" --storage ~/.karst/my-repo

Or wire it into your AI tool: MCP setup for Claude, Cursor, VS Code & more →