🪴 Brain

MCP Language Server

3 min read

MCP Language Server

isaacphi/mcp-language-server — the minimal, generic version of the LSP-as-context idea: a thin bridge that runs any stdio language server and re-exposes its capabilities as MCP tools. The agent gets IDE-grade navigation and edits derived from the language server’s native type system and symbol tables — not grep, not embeddings.

The clearest instance of “an MCP server that exposes an LSP.” See Structural Retrieval for Code for the lane.

Description

It proxies a real language server over stdio and surfaces six tools:

  • definition — “retrieves the complete source code definition of any symbol”
  • references — “locates all usages and references of a symbol throughout the codebase” (exhaustive, not top-k)
  • diagnostics — compiler/linter errors for a file
  • hover — type/signature/docs at a position
  • rename_symbol — project-wide rename
  • edit_file — line-based edits

Confirmed working with gopls (Go), rust-analyzer (Rust), pyright (Python), typescript-language-server, and clangd (C/C++) — plus “many more” servers that communicate over stdio. Because retrieval comes from the language server, it understands semantic meaning and precise callers/blast-radius that embeddings can’t.

Caveat on wording: “retrieval” here is really navigation/lookup, and rename/edit are mutations — this is a code-intelligence tool surface, not a RAG pipeline.

Download or use

Go binary, run as an MCP server pointed at a workspace + a chosen language server command; register in the agent’s MCP config (Claude Code, Cursor, etc.). One server instance per language server.

Reasoning for

This is the lowest-friction way to give a Claude Code agent type-accurate structure without adopting a whole graph engine: if the language already has a good LSP, wrap it and the agent can ask “definition of X”, “every reference to Y”, “does this file compile”. Thinner and more composable than Serena — fewer tools, no refactoring DSL — which makes it a good building block when I want one language’s intelligence rather than a 40-language toolkit.

For JS/TS work, wrap typescript-language-server and the agent gets type-accurate TS navigation with almost no setup — the sweet spot. (C#/Roslyn would wrap too in principle, but it’s untested and not the focus.)

Alternatives considered

  • Serena — richer LSP toolkit with symbol-level refactoring and 40+ languages out of the box; heavier.
  • Codebase Memory MCP — persistent tree-sitter + Hybrid-LSP graph with Cypher/trace_path; a stored graph vs live LSP calls.
  • Graphify · Understand Anything — tree-sitter syntactic graphs, no type resolution.
  • Plain agentic grep — index-free, cheaper, but false positives and no type info.

đź“– Further reading

Resources

Template: tool

Graph View

Backlinks