repo-context-mcp-server

MCP server that provides semantic code search, documentation search, and memory search for any repository. Uses LanceDB for vector storage and supports multiple embedding providers.

Tools

• search_code — Search source code for implementations, patterns, and files. Returns relevant chunks with file paths and exports.
• search_docs — Search project documentation for architecture, standards, and guidelines.
• search_memory — Search past decisions, feedback, and project context from Claude Code memory.

Prerequisites

Install Bun:

curl -fsSL https://bun.sh/install | bash

Install Ollama and pull the embedding model:

# macOS
brew install ollama

# Start Ollama (runs on http://localhost:11434)
ollama serve

# Pull the embedding model
ollama pull nomic-embed-text

Install

git clone <repo-url> ~/repos/repo-context-mcp-server
cd ~/repos/repo-context-mcp-server
bun install

Configure a repository

1. Add the MCP server to your project

From your project directory:

cd /path/to/your/project
claude mcp add repo-context -s project -- bun /path/to/repo-context-mcp-server/src/index.ts

This creates a .mcp.json file in your project root.

2. Create a config file

Create repo-context.json in your project root:

{
  "embedding": {
    "type": "ollama",
    "apiKey": null,
    "baseUrl": "http://localhost:11434",
    "model": "nomic-embed-text"
  }
}

3. Add to .gitignore

.repo-context/

4. Restart Claude Code

The server connects immediately on startup. The initial index builds in the background — searches return empty results until indexing completes.

Configuration

All fields except embedding are optional:

{
  "embedding": {
    "type": "ollama",
    "apiKey": null,
    "baseUrl": "http://localhost:11434",
    "model": "nomic-embed-text"
  },
  "codePatterns": ["**/*.{ts,tsx,js,jsx}"],
  "docPatterns": ["docs/*.md", "CLAUDE.md", ".claude/rules/*.md"],
  "skipPatterns": [
    "node_modules", ".next", "/dist/", "/build/",
    ".bundle.", ".min.", "__tests__", "__mocks__"
  ]
}

| Field | Default | Description | |---|---|---| | embedding | (required) | Embedding provider configuration | | codePatterns | ["**/*.{ts,tsx,js,jsx}"] | Glob patterns for source code files | | docPatterns | ["docs/*.md", "CLAUDE.md", ".claude/rules/*.md"] | Glob patterns for documentation files | | skipPatterns | node_modules, .next, dist, build, tests, mocks | Substrings in file paths to skip | | projectRoot | process.cwd() | Override project root (auto-detected from MCP cwd) | | memoryDir | auto-detected | Override Claude Code memory directory |

Config file resolution (first found wins):

1. <cwd>/repo-context.json
2. <cwd>/.repo-context/config.json
3. <server-dir>/config.json

Embedding providers

| Provider | Model (default) | Dimensions | Config type | |---|---|---|---| | Ollama | nomic-embed-text | 768 | ollama | | OpenAI | text-embedding-3-small | 1536 | openai | | Google | text-embedding-004 | 768 | google | | Mistral | mistral-embed | 1024 | mistral | | LM Studio | (custom) | 768 | lmstudio |

How it works

• Code indexing — TypeScript AST-aware chunking splits at top-level declarations. Falls back to sliding window (200 lines, 50-line overlap) for other file types.
• Doc indexing — Splits markdown on ## heading boundaries.
• Memory indexing — Indexes Claude Code memory files with YAML frontmatter parsing.
• Deduplication — Content hashing (SHA-256) avoids re-embedding unchanged files.
• File watching — Chokidar watches for changes with 5-second debounce, re-indexes automatically.
• Vector storage — LanceDB stores embeddings locally in .repo-context/ within the project.