ContextWeave

Code-context MCP server for coding agents. It gives an agent structural, lexical and semantic retrieval over a repository, so it can pull the definition it needs instead of reading whole files.

• Structural: tree-sitter symbols, key paths and headings for 27 languages and data/doc formats (cw languages).
• Lexical + semantic: a SQLite index with full-text search and sqlite-vec embeddings, fused and packed into a token budget.
• Graph: callers, callees, imports, inheritance, tests and config edges.

Published on PyPI as contextweave-mcp. The import package is contextweave and the commands are cw and contextweave.

Install

Requires Python 3.14+.

uv tool install contextweave-mcp
cw --version

Quickstart

From the root of a project:

cw init                 # index + wire up Claude Code (default agent)
cw status               # counts, freshness, daemon and model health
cw search "where are retries handled"

cw init indexes the project and installs the agent integration: MCP server registration, hooks, and a routing block that tells the agent when to use which tool. Remove it again with cw uninstall (the index is kept; cw forget drops it).

cw init --plugin writes a Claude Code plugin directory instead of editing project files. --dry-run shows what would change.

To run the server by hand for any MCP client:

cw serve   # MCP over stdio

MCP tools

Every tool is read-only. Results are terse text; summarised or omitted items carry cw:// refs that expand resolves later, and report when the code moved or was deleted.

Models

Semantic search, chunk annotation and query gating use local models served by llama.cpp llama-server (embed, annotate) and a classifier endpoint. No model weights are bundled or downloaded by ContextWeave; you run the servers and point the config at them. Each model has its own licence: the default embed model, jina-code-embeddings, is CC-BY-NC (non-commercial). Swap models.embed for a permissively licensed model if that does not fit your use.

Example servers:

llama-server -hf jinaai/jina-code-embeddings-1.5b-GGUF:Q8_0 --embeddings --pooling last --port 8770
llama-server -hf unsloth/Qwen3.5-4B-GGUF:Q4_K_M --port 8772

Models are optional. Without them the structural and lexical paths still work; indexing reports the skipped stages as degraded and cw index fills them in once the servers are up. cw models probe times each slot on your hardware.

Configuration

Settings are layered: built-in defaults, then ~/.config/contextweave/config.toml (honours XDG_CONFIG_HOME), then .contextweave.toml in the project root. cw config show prints the effective values.

[index]
include = ["src", "tests"]
ignore = ["vendor/**"]

[models.embed.gpu]
endpoint = "http://127.0.0.1:8770"
# Set launch to have the daemon start and supervise the server itself.
launch = ""

CLI

Platforms

Linux and WSL2 are the primary targets and run the shared per-user daemon. On Windows the server runs in-process; the daemon needs Unix sockets.

Development

uv sync
uv run ruff check . && uv run ruff format --check .
uv run pyright
uv run pytest            # add -m "not daemon" off Linux

Licence

Mozilla Public License 2.0. Model weights used at runtime are not part of this project and are covered by their own licences.