large-repo-mcp-server
A stdio-based MCP server for fast, bounded search and navigation of large code repositories using ripgrep, with tools for project-wide search, symbol lookup, file listing, and code reading.
README
Large Repo MCP Server
Stdio-based Model Context Protocol server designed as a standard toolkit for large and complex repositories. Uses ripgrep for fast, bounded search across codebases of any size.
Architecture
┌─────────────────────────────────────────────────────────────────┐
│ MCP Client (Codex CLI / Claude Desktop / Claude Code) │
│ │
│ tools/call ──► JSON-RPC 2.0 request │
│ Content-Length: N\r\n\r\n{...} │
└─────────────┬───────────────────────────────────▲───────────────┘
│ stdin │ stdout
▼ │
┌─────────────────────────────────────────────────────────────────┐
│ large-repo-mcp server │
│ │
│ ┌──────────┐ ┌────────────┐ ┌──────────────────────┐ │
│ │ Frame │───►│ JSON-RPC │───►│ Tool Dispatch │ │
│ │ Parser │ │ Router │ │ │ │
│ │ │ │ │ │ project_search_rg │ │
│ │ Content- │ │ initialize │ │ symbol_search │ │
│ │ Length │ │ tools/list │ │ read_range │ │
│ │ framing │ │ tools/call │ │ list_files │ │
│ └──────────┘ │ ping │ └──────────┬───────────┘ │
│ └────────────┘ │ │
│ │ │
│ ┌────────────────────────────────────────────▼──────────┐ │
│ │ Security Layer │ │
│ │ │ │
│ │ • Path confinement (resolve + realpath + root check) │ │
│ │ • Command allowlist (rg only, shell: false) │ │
│ │ • Request size limit (1 MB) │ │
│ │ • Response size limit (200 KB) │ │
│ │ • Null-byte rejection │ │
│ └───────────────────────────┬───────────────────────────┘ │
│ │ │
│ ▼ │
│ ┌─────────────────┐ │
│ │ rg (ripgrep) │ │
│ │ subprocess │ │
│ │ │ │
│ │ • 15s timeout │ │
│ │ • JSON output │ │
│ │ • Streaming │ │
│ └─────────────────┘ │
│ stderr ──► │
│ structured JSON logs │
└─────────────────────────────────────────────────────────────────┘
Request Flow
Client Server ripgrep
│ │ │
│ Content-Length: N\r\n\r\n │ │
│ {"jsonrpc":"2.0", │ │
│ "method":"tools/call",...} │ │
│──────────────────────────────►│ │
│ │ validate request size (≤1MB) │
│ │ parse JSON-RPC frame │
│ │ validate tool + arguments │
│ │ resolve & confine paths │
│ │ │
│ │ spawn rg --json ... │
│ │──────────────────────────────►│
│ │ │
│ │ stream results (line-by-line)│
│ │◄──────────────────────────────│
│ │ │
│ │ enforce match/byte limits │
│ │ kill rg if limit reached │
│ │ │
│ Content-Length: M\r\n\r\n │ │
│ {"jsonrpc":"2.0", │ │
│ "result":{...}} │ │
│◄──────────────────────────────│ │
│ │ │
Safety Guarantees
| Guarantee | Mechanism | Limit |
|---|---|---|
| Fast, bounded search | ripgrep with match caps | 500 max matches |
| Response size cap | Byte-level output budget | 200 KB |
| Subprocess timeout | setTimeout + child.kill() |
15 seconds |
| Path confinement | path.resolve + realpath + root prefix check |
repo root only |
| Symlink escape prevention | fs.realpath() resolves then re-validates |
double-checked |
| Command execution | Allowlist (rg only) + shell: false |
no shell injection |
| Request size limit | Frame-level rejection before parse | 1 MB |
| Input sanitization | Null-byte rejection, type validation | all tool inputs |
| Child process cleanup | Tracked set, killed on shutdown signals | SIGTERM/SIGINT/exit |
Supported Clients
- Codex CLI
- Claude Desktop
- Claude Code
- Any MCP client supporting stdio transport (protocol version
2024-11-05)
Requirements
- Node.js 18+
- ripgrep (
rg) onPATHsymbol_searchrequires PCRE2 support (most official packages include it)
Verify ripgrep:
rg --version # should show version
rg --pcre2-version # should show PCRE2 version
Install ripgrep:
| Platform | Command |
|---|---|
| macOS | brew install ripgrep |
| Ubuntu/Debian | sudo apt-get install ripgrep |
| Fedora | sudo dnf install ripgrep |
| Windows | winget install BurntSushi.ripgrep.MSVC |
| Cargo | cargo install ripgrep --features pcre2 |
Install
npm install
npm run build
Run
npm start
Default repo root behavior:
- If
REPO_ROOTis unset, repo root is current working directory (process.cwd()). - Set
REPO_ROOTto override explicitly.
PowerShell:
$env:REPO_ROOT = "C:\absolute\path\to\repo"
npm start
MCP Configuration
Codex (Global)
Add to ~/.codex/config.toml:
[mcp_servers.large_repo_mcp]
command = "node"
args = ["C:/absolute/path/to/mcp/dist/server.js"]
startup_timeout_sec = 30
[mcp_servers.large_repo_mcp.env]
REPO_ROOT = "."
Codex (Project-Local)
Add to .codex/config.toml in a project:
[mcp_servers.large_repo_mcp]
command = "node"
args = ["C:/absolute/path/to/mcp/dist/server.js"]
startup_timeout_sec = 30
[mcp_servers.large_repo_mcp.env]
REPO_ROOT = "."
Claude Desktop
Add to Claude Desktop MCP config:
{
"mcpServers": {
"large_repo_mcp": {
"command": "node",
"args": ["C:/absolute/path/to/mcp/dist/server.js"],
"env": {
"REPO_ROOT": "C:/absolute/path/to/target-repo"
}
}
}
}
Claude Code
Add to .mcp.json in a project or ~/.claude/mcp.json globally:
{
"mcpServers": {
"large_repo_mcp": {
"command": "node",
"args": ["C:/absolute/path/to/mcp/dist/server.js"],
"env": {
"REPO_ROOT": "."
}
}
}
}
Environment Variables
| Variable | Default | Description |
|---|---|---|
REPO_ROOT |
process.cwd() |
Repository root directory |
LARGE_REPO_MCP_LOG_LEVEL |
error |
Log level: error, warn, info, debug |
LARGE_REPO_MCP_DEBUG |
— | Set to 1 to force debug logging |
Tools
project_search_rg
Search the repository with ripgrep. Supports regex patterns and optional glob filters.
┌──────────────────────────────────────────────────────┐
│ project_search_rg │
│ │
│ Input Output │
│ ───── ────── │
│ pattern (string, required) matches[] │
│ globs (string[], max 50) .path │
│ maxMatches (1-500, def 100) .line │
│ .column │
│ .text │
│ .submatches[] │
│ truncated │
│ truncateReason │
│ timedOut │
│ serverVersion │
│ durationMs │
└──────────────────────────────────────────────────────┘
Example call:
{
"name": "project_search_rg",
"arguments": {
"pattern": "TODO|FIXME",
"globs": ["*.ts", "*.tsx"],
"maxMatches": 50
}
}
symbol_search
Search for an exact symbol using word-boundary matching (\b...\b). Auto-detects project type (TypeScript or Python) and scopes file types accordingly.
┌──────────────────────────────────────────────────────┐
│ symbol_search │
│ │
│ Input Output │
│ ───── ────── │
│ symbol (string, required) projectType │
│ maxMatches (1-500, def 100) matchMode │
│ matches[] │
│ .path │
│ Detection logic: .line │
│ ┌────────────────────┐ .column │
│ │ tsconfig.json? │ .text │
│ │ yes → typescript │ .submatches[] │
│ │ no → check files │ │
│ │ .py only → python│ │
│ │ .ts only → ts │ │
│ │ fallback → ts │ │
│ └────────────────────┘ │
└──────────────────────────────────────────────────────┘
Example call:
{
"name": "symbol_search",
"arguments": {
"symbol": "handleRequest",
"maxMatches": 20
}
}
read_range
Read a specific line range from a file. Path must be relative and resolve inside the repo root.
┌──────────────────────────────────────────────────────┐
│ read_range │
│ │
│ Input Output │
│ ───── ────── │
│ path (relative, required) path (normalized) │
│ startLine (int, required) requestedLines │
│ endLine (int, required) returnedLines │
│ lines[] │
│ Constraints: .line │
│ • max 500 lines per call .text │
│ • path confined to repo root truncated │
│ • symlinks resolved + checked truncateReason │
└──────────────────────────────────────────────────────┘
Example call:
{
"name": "read_range",
"arguments": {
"path": "src/server.ts",
"startLine": 1,
"endLine": 50
}
}
list_files
List repository files using ripgrep --files with optional glob filters. Excludes .git, node_modules, dist, build, and coverage by default.
┌──────────────────────────────────────────────────────┐
│ list_files │
│ │
│ Input Output │
│ ───── ────── │
│ globs (string[], max 50) files[] (paths) │
│ maxResults (1-500, def 500) returned │
│ truncated │
│ Auto-excluded dirs: truncateReason │
│ .git, node_modules, dist, timedOut │
│ build, coverage │
└──────────────────────────────────────────────────────┘
Example call:
{
"name": "list_files",
"arguments": {
"globs": ["src/**/*.ts"],
"maxResults": 100
}
}
Project Type Detection
The server auto-detects whether a repository is TypeScript or Python to scope symbol_search file types. Detection runs once per process and is cached.
┌─────────────────────┐
│ tsconfig.json │
│ exists? │
└──────┬──────────────┘
yes │ no
┌──────────┘ │
▼ ▼
┌──────────┐ ┌────────────────────┐
│TYPESCRIPT│ │ Python project │
└──────────┘ │ file exists? │
│ (pyproject.toml, │
│ requirements.txt, │
│ setup.py, etc.) │
└──────┬─────────────┘
yes │ no
┌──────────┤ │
▼ │ ▼
┌────────────┐ │ ┌──────────────────┐
│package.json│ │ │ package.json │
│ exists? │ │ │ exists? │
└─────┬──────┘ │ └──────┬───────────┘
yes │ no │ yes │ no
│ │ │ │ │ │ │
▼ │ ▼ │ ▼ │ ▼
┌────────┐│ ┌──────┐ │ ┌──────────┐ ┌──────────────┐
│ Both ││ │PYTHON│ │ │TYPESCRIPT│ │ Count .ts vs │
│ scan ││ └──────┘ │ └──────────┘ │ .py files │
│ files │▼ ▼ └──────┬───────┘
└───┬────┘ ┌──────────┐ more │ more
│ │ No pkg, │ .py │ .ts
▼ │ has py │ │ │
┌────────────┐ │ file → │ ▼ ▼
│ .py only → │ │ PYTHON │ ┌──────┐ ┌──┐
│ PYTHON │ └──────────┘ │PYTHON│ │TS│
│ .ts only → │ └──────┘ └──┘
│ TS │
│ fallback → │
│ TS │
└────────────┘
Dev Workflow
npm run typecheck # type-check without emitting
npm test # build + run unit & integration tests
npm run lint # run ESLint
npm run format:check # check Prettier formatting
Troubleshooting
| Problem | Solution |
|---|---|
rg not found |
Install ripgrep and ensure it is on PATH |
symbol_search PCRE2 error |
Install ripgrep with PCRE2 support (most official packages include it) |
Request exceeds 1048576 bytes |
Split large requests; max inbound JSON-RPC frame body is 1 MB |
truncated: true |
Increase maxMatches/maxResults (up to bounds), narrow your query, or add globs |
No results from symbol_search |
Check projectType in response — detection may have picked the wrong language; use project_search_rg as a fallback |
License
MIT
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。