code-context

Status: stable (v1.0.0). A Python MCP server with local RAG for Claude Code. Implements the code-context Tool Protocol v1.2 defined by context-template.

What it does

When you point Claude Code at a repo, you give it CLAUDE.md for static context. code-context adds dynamic context via 7 MCP tools:

• search_repo(query, top_k?, scope?) — hybrid retrieval across the codebase: vector embeddings (semantic) fused with BM25 keyword search (exact identifiers) via Reciprocal Rank Fusion. Optional cross-encoder reranking (off by default — enable with CC_RERANK=on).
• recent_changes(since?, paths?, max?) — recent git commits, optionally filtered.
• get_summary(scope?, path?) — structured project summary (name, stack, key modules, stats).
• find_definition(name, language?, max?) — locate where a symbol (function, class, method, type) is defined. Use INSTEAD of Grep for def X / class X / function X patterns. Returns repo-relative paths with line ranges and the symbol's kind (function, class, method, interface, struct, enum, record). For Markdown files, also finds doc sections by heading text (kind = section).
• find_references(name, max?) — list every line mentioning a named symbol. Use INSTEAD of grep -n "X" when the user asks "who calls X?" or "where is X used?". Word-boundary matched, so log doesn't return logger.
• get_file_tree(path?, max_depth?, include_hidden?) — repo-relative directory tree, gitignore-aware. Use INSTEAD of Bash: ls -R or Bash: tree for orientation prompts ("show me the project structure", "what's in this module?"). Returns hierarchical FileTreeNode with file sizes; honors .gitignore; defaults to depth 4.
• explain_diff(ref, max_chunks?) — AST-aligned chunks affected by the diff at ref (full SHA, HEAD, HEAD~N, branch). Use INSTEAD of Bash: git show <sha> for "what does this commit do" questions. The chunker resolves which whole functions/classes were touched, not raw line additions.

Architecture: hexagonal (ports & adapters). 9 driven ports with default implementations (sentence-transformers embeddings, NumPy+Parquet vector store, tree-sitter / line chunker, filesystem code source, git CLI, filesystem introspector, SQLite FTS5 keyword index, cross-encoder reranker, SQLite-backed symbol index). All swappable.

Install

pip install code-context-mcp
# or, if you don't want torch (~2 GB), use the OpenAI embeddings backend:
pip install code-context-mcp[openai]

The PyPI distribution is code-context-mcp (the unhyphenated code-context name was squatted by an unrelated, abandoned project from 2023; see CHANGELOG for context). The Python module is still code_context and the CLI binaries are still code-context and code-context-server, so quickstart commands and from code_context import ... are unchanged.

Note: the default install pulls sentence-transformers + the all-MiniLM-L6-v2 model on first run. Plan for ~2 GB of disk after first reindex (torch ≈ 2 GB, model ≈ 90 MB). Use the [openai] extra to avoid torch entirely.

Quickstart

cd /path/to/your/repo
claude mcp add code-context --command code-context-server
# Open Claude Code. From v0.9.0 the server starts in <1 s on a previously-indexed
# repo; the first reindex (and any subsequent ones) run on a background thread,
# so queries are never blocked. Cold start: queries return [] until the first
# bg reindex completes (~30-60 s on a typical repo with all-MiniLM on CPU).
# Edit-cycle reindex is sub-10 s thanks to v0.8.0's dirty_set tracking.

Live mode (optional)

If you want every save in the repo to flow into the index without manual code-context reindex:

pip install code-context-mcp[watch]   # adds watchdog
export CC_WATCH=on
claude mcp add code-context --command code-context-server

Edits are debounced for ~1 s (configurable via CC_WATCH_DEBOUNCE_MS) and then trigger a background reindex. Default off — opt-in.

For OpenAI embeddings:

export CC_EMBEDDINGS=openai
export OPENAI_API_KEY=sk-...
claude mcp add code-context --command code-context-server

GPU support

code-context auto-detects the best available device for embeddings and cross-encoder rerank:

• CUDA: install torch with the CUDA wheels (pip install torch --index-url https://download.pytorch.org/whl/cu121). The first query after a cold start will use GPU automatically. Expect cross-encoder p50 ≤ 100 ms on most consumer GPUs.
• Apple Silicon (MPS): detected automatically on macOS with M-series chips. Some sentence-transformers operations are not yet stable on MPS; if the model fails to load, code-context logs a warning and falls back to CPU.
• CPU: the default fallback. With v1.5's distilled cross-encoder (MiniLM-L-2-v2), hybrid rerank p50 is ~1.1 s on CPU — usable interactively from Claude Code.

No env var or config flag is required.

Windows: Microsoft Store Python sandbox

If you installed Python from the Microsoft Store (the default in some Windows SKUs), the OS silently redirects writes from %LOCALAPPDATA% (where platformdirs places the default cache) to a per-app sandbox under:

%LOCALAPPDATA%\Packages\PythonSoftwareFoundation.Python.3.X_qbz5n2kfra8p0\LocalCache\Local\code-context\

This is fine — the index works — but code-context reports the nominal cache path, not the sandboxed real path. If you can't find the cache where code-context status prints, look under Packages\...\LocalCache\... or set CC_CACHE_DIR explicitly to a path outside the sandbox:

$env:CC_CACHE_DIR = "C:\Users\<you>\code-context-cache"

To avoid the sandbox entirely, install Python from python.org instead of the Microsoft Store.

Making Claude actually use these tools

Claude Code defaults to its built-in tools (Bash, Grep, Glob, Read) over MCP servers because it knows them best. To get the value of code-context, give Claude an explicit hint by adding a section like this to your project's CLAUDE.md:

## Context tools

This repo has the [code-context](https://github.com/nachogeinfor-ops/code-context) MCP server installed. Prefer it over built-in tools:

- **`search_repo(query, top_k?, scope?)`** — for conceptual questions like "where do we handle authentication" or "how is caching implemented". Use this instead of `Grep` whenever the query isn't an exact string match.
- **`recent_changes(since?, paths?, max?)`** — for "what changed recently" / commit-history questions. Use this instead of shelling out to `git log`.
- **`get_summary(scope?, path?)`** — for project orientation at session start, or to inspect a specific module.
- **`find_definition(name, language?, max?)`** — for "where is X defined?". Use this instead of `Grep` for `def X` / `class X` patterns; tree-sitter-indexed at reindex time, so it's faster and more accurate than scanning text.
- **`find_references(name, max?)`** — for "who calls X?" / "where is X used?". Use this instead of `grep -n`; word-boundary matched so `log` won't match `logger`.
- **`get_file_tree(path?, max_depth?, include_hidden?)`** — for "show me the project structure" / "what's in this module?". Use this instead of `Bash: ls -R` / `Bash: tree`; gitignore-aware and structured (file sizes included).
- **`explain_diff(ref, max_chunks?)`** — for "what does this commit do?" / "what changed in HEAD~3?". Use this instead of `Bash: git show <sha>`; the chunker resolves whole functions/classes that were touched, not raw line additions.

Without this hint, Claude will work fine — it just won't reach for the MCP tools, which means the index goes unused. The hint is one paragraph; copy-paste it.

CLI

code-context-server is the MCP binary; you don't run it directly. The companion code-context CLI helps administer the index:

code-context status                                       # print index health + dirty/deleted counts
code-context doctor                                       # run env + index health checks (no side effects)
code-context reindex                                      # incremental by default (only changed files)
code-context reindex --force                              # full reindex (post-model-upgrade or cache reset)
code-context query "where do we validate user emails"     # debug, no MCP
code-context clear --yes                                  # delete the cache for this repo
code-context refresh                                      # trigger a reindex + wait for swap (since v1.10.0)
code-context cache export --output cache.tar.gz           # bundle the active index (since v1.10.0)
code-context cache import cache.tar.gz                    # restore a bundle (rejects version mismatches; --force overrides)

doctor is the first stop when something looks wrong — it surfaces missing dependencies, an unwritable cache, an absent HF model cache, a corrupted index, etc., without doing anything destructive. Exit code is 0 if every check passed, 1 if anything failed.

Configuration

Configured via env vars. See docs/configuration.md for the full list. Most-used:

Var	Default
CC_EMBEDDINGS	local (or openai)
CC_EMBEDDINGS_MODEL	all-MiniLM-L6-v2
CC_INCLUDE_EXTENSIONS	.py,.js,.ts,.jsx,.tsx,.go,.rs,.java,.c,.cpp,.h,.hpp,.md,.yaml,.yml,.json
CC_CHUNKER	treesitter (AST-aware for 9 languages: Python, JavaScript, TypeScript, Go, Rust, C#, Java, C++, Markdown — line fallback for the rest) — set line for v0.1.x behavior
CC_CACHE_DIR	platformdirs user cache
CC_TELEMETRY	off (opt-in; see below)

Switching embeddings backend

Since v2.0.6, the default embeddings backend is onnxruntime. The torch backend remains available for users who need CUDA acceleration or want to use a model not on the ONNX-supported list.

# Default (fast cold start, CPU-only):
export CC_EMBEDDINGS_BACKEND=onnx   # implicit default

# Opt back to torch (slower cold start, supports CUDA/MPS):
export CC_EMBEDDINGS_BACKEND=torch

Models with verified ONNX exports in v2.0.6:

• Embeddings: all-MiniLM-L6-v2 (default)
• Reranker: cross-encoder/ms-marco-MiniLM-L-2-v2 (default)

Other registered models (BAAI/bge-base-en-v1.5, jinaai/jina-embeddings-v2-base-code, nomic-ai/CodeRankEmbed) automatically fall back to torch with an info log when selected under the ONNX backend. code-context doctor shows the active backend.

Telemetry (opt-in)

Telemetry is off by default and always opt-in. On your first run against a new repo, the CLI (code-context query/reindex/status) asks once whether to enable it; your answer is persisted in the per-repo cache and respected on subsequent runs. Non-interactive callers (piped CLI, MCP stdio server) never prompt and default to off — set CC_TELEMETRY=on explicitly to opt in for those.

What's collected when enabled: a weekly heartbeat and session aggregates to PostHog Cloud. Never PII, query text, code content, repo paths, file names, or IPs. See docs/telemetry.md for the full schema, what's not collected, and how the anonymous install ID is derived.

CC_TELEMETRY env var always overrides the per-repo marker.

Documentation

• Public API (v1) — what's stable; what's not. Read this before depending on code-context from another project.
• Configuration — every env var with examples (chunker strategies, hybrid search, symbol index, background reindex, watch mode, …).
• Architecture — hexagonal diagram, port contracts, indexing lifecycle, Sprint 7 background-thread + bus.
• Eval suite — NDCG@10 / MRR / latency baselines per retrieval mode.
• Releasing — Trusted Publisher setup, per-release checklist.
• Extending — write your own embeddings provider, vector store, or chunker.

Status

v1.0.0 — stable. Public surface frozen; v1.x will only add. See docs/v1-api.md for the commitment scope and CHANGELOG.md for what shipped in each version.

License

MIT.