lsp-mcp-server
Exposes type-aware code navigation and fast file search to AI agents via language servers, enabling definitions, references, symbols, and file lookup without reading entire codebases.
README
lsp-mcp-server
A lightweight MCP server that exposes type-aware code navigation and fast file search to AI agents via language servers. It manages language server processes on demand, routes LSP requests from MCP tool calls, and returns structured results — letting agents jump to definitions, find references, list symbols, and locate files without reading entire codebases.
Installation
Option 1: Claude Code Plugin (recommended)
One command to clone, install, and register as a Claude Code plugin:
git clone https://github.com/garretpremo/lsp-mcp-server.git ~/.local/share/lsp-mcp-server \
&& cd ~/.local/share/lsp-mcp-server && bun install \
&& bash scripts/install-plugin.sh
Restart Claude Code (or run /reload-plugins) to activate. The plugin adds a skill that guides Claude on when to use each tool, plus the MCP server runs automatically for each project.
Option 2: Per-session (no install)
git clone https://github.com/garretpremo/lsp-mcp-server.git ~/tools/lsp-mcp-server \
&& cd ~/tools/lsp-mcp-server && bun install
Then start Claude Code with:
claude --plugin-dir ~/tools/lsp-mcp-server/plugin
Option 3: Manual MCP config
Add to your project's .mcp.json or ~/.claude/settings.json:
{
"mcpServers": {
"lsp-mcp-server": {
"command": "bun",
"args": ["run", "/path/to/lsp-mcp-server/src/index.ts", "--project", "."]
}
}
}
Option 4: Standalone (any MCP client)
bun run src/index.ts --project /path/to/your/project
Communicates via stdio JSON-RPC — works with any MCP-compatible client.
Available Tools
| Tool | Description |
|---|---|
find_file |
Fuzzy file name search against the project file index |
reindex |
Rebuild the file index (full project or a subdirectory) |
go_to_definition |
Jump to the definition of a symbol at a given file position |
find_references |
Find all references to a symbol at a given file position |
document_symbols |
List all symbols (classes, functions, etc.) in a file |
workspace_symbol |
Search for symbols by name across the entire project |
All LSP navigation tools accept optional enrichment flags: includeKind, includeContainer, and includeDocstring.
Configuration
Place a config.json in your project root to override language server commands or the request timeout:
{
"requestTimeout": 15000,
"languageServers": {
"typescript": {
"command": "typescript-language-server",
"args": ["--stdio"]
},
"java": {
"command": "jdtls",
"args": []
}
}
}
If no config.json is found, defaults are used. requestTimeout is in milliseconds (default: 10000).
Benchmarks: lsp-mcp-server vs JetBrains MCP
Tested against a 4,381-file Angular/TypeScript project (Nx monorepo).
File Search — "document-viewer"
| lsp-mcp-server | JetBrains MCP | |
|---|---|---|
| Response size | ~1,500 chars | 9,792 chars |
| Results | 10 files, ranked by relevance | 40 items (files + dirs), unranked, duplicated by worktrees |
| Format | Absolute + relative paths | Flat path list with worktree duplicates |
Document Symbols — reports.service.ts
| lsp-mcp-server | JetBrains MCP | |
|---|---|---|
| Response size | ~5,200 chars | 4,486 chars |
| Result type | 31 structured symbols (name, kind, container, snippet) | Raw file text (no symbols tool available) |
| Structured | Yes — class, method, property with hierarchy | No — LLM must parse symbols from raw text |
Go To Definition — HttpClient import
| lsp-mcp-server | JetBrains MCP | |
|---|---|---|
| Response size | ~550 chars | 34,868 chars |
| Result type | 1 precise definition location | 50 text matches (grep-style, not semantic) |
| Semantic | Yes — LSP-powered, type-aware | No — text search with search_in_files_by_text |
Agent Token Usage (identical tasks, Sonnet model)
| lsp-mcp-server | JetBrains MCP | |
|---|---|---|
| Total tokens | 18,910 | 25,994 |
| Tool calls | 3 | 10 |
Summary
- 6.5x smaller file search responses
- 63x smaller definition lookup responses
- 27% fewer tokens consumed by the agent
- Structured symbol data vs raw text
- True LSP semantics vs text-based grep
Prerequisites
The following language servers must be installed and available on your PATH:
- TypeScript / JavaScript:
typescript-language-servernpm install -g typescript-language-server typescript - Java:
jdtls(Eclipse JDT Language Server)
Language servers are started lazily — only when a file of the matching type is first accessed.
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。