@codeweave/mcp

@codeweave/mcp

Give your AI agent structured code understanding — not just file dumps.

CodeWeave is an MCP server that gives AI agents cheap, precise code intelligence. Instead of dumping entire files into context, your agent queries local indexes — AST, call graph, type graph, hybrid semantic search — and gets back only what it needs.

Less tokens. More relevant context. Better decisions.

The semantic search pipeline is the heart of the system: a 6-stage hybrid engine combining vector embeddings, full-text search, and structural density scoring. Tested extensively across large production codebases — Java monoliths, TypeScript monorepos, Python ML pipelines, Go microservices — with consistently strong retrieval accuracy.

Actively developed. New tools and improvements ship regularly. Contributions and feedback are welcome.

Quick Start

cd your-project
npx @codeweave/mcp

That's it. The setup wizard handles everything:

• Installs @codeweave/mcp globally
• Installs Ollama if needed
• Downloads the embedding model
• Configures your MCP client (Claude Code, VS Code)
• Indexes your project

Note: The first run requires a one-time download of Ollama and the embedding model. This takes a few minutes but only happens once.

Open your project in Claude Code or VS Code and start asking questions.

Tools

8 tools organized around the code understanding workflow:

Discover

Tool	Purpose
semantic_search	Search by meaning — finds functions even when you don't know exact names. Hybrid vector + keyword search with density-based reranking.
get_module_summary	Browse a directory's functions and classes with signatures. Auto-adapts detail level to module size.

Read

Tool	Purpose
get_function_source	Get a specific function's source code — no need to read entire files. Supports class context and surrounding lines.

Analyze

Tool	Purpose
get_dependencies	What does this function call? Cross-validates AST with docstring @deps. Categorizes: confirmed, AST-only, docstring-only, unresolved.
get_impact_analysis	Blast radius of a change. Combines call graph + type graph. Risk levels: high (direct callers), medium (indirect), low (transitive).
get_stale_docstrings	Find missing or outdated docstrings. Detects @deps drift, missing @tags, and undocumented functions.

Maintain

Tool	Purpose
reindex	Manually trigger index update. Usually unnecessary — file watcher auto-reindexes on changes.
get_index_status	Index health dashboard: file/function counts, embedding status, call graph stats, language breakdown.

How It Works

Source Code
    │
    ▼
tree-sitter AST  ───>  Function Index (in-memory)
                              │
                   ┌──────────┼──────────┐
                   ▼          ▼          ▼
              Call Graph  Type Graph  Embeddings
              (JSON)      (JSON)     (LanceDB)
                   │          │          │
                   └──────────┼──────────┘
                              ▼
                       8 MCP Tools  ───>  AI Agent

• Parse — tree-sitter extracts every function, class, method, and interface across 7 languages
• Embed — Qwen3-Embedding-0.6B generates vector embeddings for semantic search
• Index — LanceDB stores vectors with BM25 full-text index alongside
• Graph — Call graph tracks who-calls-whom with type-aware resolution; type graph tracks inheritance and implementations
• Watch — File watcher detects changes and incrementally reindexes affected files
• Serve — 8 tools exposed over MCP protocol (stdio), ready before indexing completes

Semantic Search

The search pipeline is where CodeWeave really shines. It's not just vector similarity — it's a multi-stage system designed to surface the most relevant and important code:

6-Stage Pipeline:

• Exact name match — Fast path for known function names (score 0.95+)
• Vector search — Embed the query, find semantically similar functions (over-fetches 3x for reranking headroom)
• Full-text search — BM25 keyword matching catches what embeddings miss
• RRF merge — Reciprocal Rank Fusion combines both result lists without needing score calibration
• Exact match boost — Functions whose name matches the query get priority
• Density reranking — Structural signals determine information density, pushing trivial code down

Density Scoring uses 7 language-agnostic structural signals:

Signal	What it measures
Body size	Larger functions carry more behavior (log-scaled)
Docstring presence	Documented code is more likely to be important
Docstring richness	Tags, deps, side effects indicate well-maintained code
Parameter count	More params = more complex behavior
Call graph centrality	Functions called by many others are architectural anchors
Visibility	Public > protected > private
Kind	Classes > methods/functions > interfaces

Penalties prevent noise from dominating results:

• Accessors (getters/setters) — pure data access, no behavior
• Constructors — many params inflate scores, but they're just assignments
• Test files — large bodies don't mean important behavior (unless you're searching for tests)

Graceful degradation: If Ollama is unavailable, search falls back to full-text only. AST-based tools (dependencies, impact analysis, module summary) work without any embedding infrastructure.

Why These Technologies

Every technology choice serves the core goal: local, fast, zero-config code understanding.

Technology	Why
tree-sitter	One parsing framework for all 7 languages. Mature, fast, battle-tested. Gives us full AST access without writing 7 different parsers from scratch.
LanceDB	Embedded vector database — no external server, no Docker, no configuration. Just a directory on disk. Supports both vector search and BM25 full-text search in a single engine.
Qwen3-Embedding-0.6B	The secret weapon. Just 0.6B parameters but delivers embedding quality that rivals models 10x its size for code understanding. Tested across large production codebases — Java enterprise monoliths, TypeScript monorepos, Python data pipelines — with consistently excellent retrieval accuracy. Runs locally via Ollama, fast enough for real-time reindexing, lightweight enough for any developer machine.
RRF (Reciprocal Rank Fusion)	Proven technique from information retrieval research. Merges ranked lists from different scoring systems (vector similarity vs. BM25 relevance) without needing score calibration. Simple, robust, effective.
MCP Protocol	Standard interface for AI tool integration. One server works with Claude Code, VS Code, Cursor, and any MCP-compatible client.

Supported Languages

Language	Functions	Calls	Imports	Types	Test Detection
Python	functions, methods, classes	call sites	import/from-import	class inheritance, type hints	pytest, unittest
TypeScript	functions, arrows, methods, classes, interfaces	call sites	named/default/namespace imports	implements, extends, member types	jest, vitest, playwright
JavaScript	(same as TypeScript)	(same as TypeScript)	(same as TypeScript)	(same as TypeScript)	jest, vitest, mocha
Go	functions, methods (receiver), structs	call sites	import specs	implicit interfaces, structs	testing, testify
Rust	functions, methods (impl), structs, enums	call sites	use declarations	impl Trait for Type	#[test], #[cfg(test)]
Java	methods, constructors, classes, interfaces	method invocations	import declarations	extends, implements	JUnit, Mockito, AssertJ
C#	methods, constructors, classes, structs, interfaces, records	invocations	using directives	base types, interface impl	NUnit, xUnit, Moq

Every language parser also provides:

• Noise filtering — built-in lists of standard library calls (e.g., console.log, fmt.Println, System.out.println) that get filtered from dependency analysis
• Structural hints — AST-confirmed classifications (constructor, abstract, getter/setter, test) that feed into density scoring

Configuration

CodeWeave works zero-config out of the box. For customization, create .code-context/config.yaml:

embedding:
  model: "qwen3-embedding:0.6b"     # Embedding model name
  ollamaUrl: "http://localhost:11434" # Ollama API endpoint
  dimensions: 1024                    # Vector dimensions
  batchSize: 50                       # Embedding batch size

parser:
  sourceRoot: "src"                   # Strip this prefix from module paths
  ignore:
    - "**/*.generated.*"              # Additional ignore patterns
    - "**/vendor/**"

search:
  rrfK: 60                           # RRF smoothing constant
  expandCamelCase: true               # Expand camelCase in search chunks
  density:
    enabled: true                     # Density-based reranking
    accessorPenalty: 0.6              # Penalty for getters/setters
    constructorPenalty: 0.7           # Penalty for constructors
    testFilePenalty: 0.5              # Penalty for test files

watcher:
  debounceMs: 500                     # File change debounce
  minIntervalMs: 2000                 # Minimum reindex interval

indexing:
  maxFileSizeKb: 500                  # Skip files larger than this

CLI Tools

# Full project initialization (AST + embeddings + graphs)
codeweave-init [path] [--force] [--no-embed]

# Incremental reindex (only changed files)
codeweave-reindex [--all] [--files=path1,path2] [--stdin]

# Docstring coverage report
codeweave-check-docstrings [--strict] [files...]

Manual Setup

If you prefer step-by-step instead of npx @codeweave/mcp:

# 1. Install globally
npm install -g @codeweave/mcp

# 2. Install Ollama and pull the embedding model
# macOS
brew install ollama
# Linux
curl -fsSL https://ollama.com/install.sh | sh

ollama pull qwen3-embedding:0.6b

# 3. Index your project
cd your-project
codeweave-init

# 4. Configure your MCP client

Claude Code — add .mcp.json to your project root:

{
  "mcpServers": {
    "codeweave": {
      "command": "codeweave-server"
    }
  }
}

VS Code — add .vscode/mcp.json:

{
  "servers": {
    "codeweave": {
      "command": "codeweave-server"
    }
  }
}

Monorepo Support

CodeWeave auto-detects workspaces in monorepos by scanning for manifest files (package.json, build.gradle, pom.xml, go.mod, Cargo.toml, pyproject.toml, etc.):

my-project/
├── backend/build.gradle    → workspace "backend"
├── mobile/package.json     → workspace "mobile"
└── shared/package.json     → workspace "shared"

Each workspace gets its own isolated index, call graph, type graph, and vector store. Tools accept an optional workspace parameter — omit it to search across all workspaces.

Git Worktree Support

CodeWeave automatically detects git worktrees (including Claude Code's /worktree). On first start in a worktree, it copies the main repo's cache for a fast warm start (~2s instead of 30s+). After that, each worktree maintains its own fully isolated index.

• Automatic — no configuration needed
• Isolated — worktree changes don't affect the main repo's cache
• Incremental — only files that differ from the main branch are re-parsed and re-embedded

Requirements

• Node.js 20+
• Ollama — for semantic search embeddings. Install via the setup wizard or manually from ollama.com. AST-based tools work without Ollama.

Status

CodeWeave is under active development. The core indexing pipeline and all 8 tools are stable and tested across production codebases in all 7 supported languages.

Feedback, bug reports, and contributions are welcome — open an issue at github.com/semihkayan/codeweave-mcp.

License

Apache 2.0