TypeScript AST MCP Server

A Model Context Protocol (MCP) server for deep structural analysis of TypeScript and JavaScript source code. Unlike text-based search, this server parses your code with the TypeScript Compiler API and walks the real AST - types, functions, call relationships, interface satisfaction, and more.

Most tools run on the syntactic tier (ts.createSourceFile, no type resolution) because it is fast and needs no build. find_implementations runs on the semantic tier - a full ts.Program and its type checker - because deciding whether a class satisfies an interface cannot be done from syntax alone. Each tool's tier is stated in the table below.

Companion to py-ast-mcp, the same idea applied to Python; a Go AST MCP server by another author served as prior art for this one.

Features

• 20 analysis tools covering file-level, directory-level, and cross-file structural queries

• Fast syntactic parsing - ts.createSourceFile() per file, under 5ms; cross-file tools re-parse in scope, so results always reflect the latest edit

• Arrow function awareness - const foo = () => {} treated as first-class functions throughout

• Signature extraction - parameters, class-qualified names, and return types as written. These come from the annotation, so an unannotated return reads as blank rather than as the inferred type

• Call graph generation with Mermaid diagrams, forward and reverse traversal, file or package scope

• Cyclomatic complexity computation per function

• Interface implementation discovery - explicit implements plus structural matches decided by the type checker's own assignability rules, so member types and call signatures count, not just member names

• Structural diffing between file versions (added/removed/modified symbols)

• Cursor-position awareness for IDE integrations

• JSX/TSX parsing - .tsx/.jsx files parse under the right ScriptKind, so every tool works on React source. There is no React-specific analysis: a component is a function like any other

• TypeScript-specific quality checks - any casts, non-null assertions, floating promises, empty catches, double assertions

• Dead code detection - unexported symbols never referenced within their own file. Unexported means file-local, so that is the whole search space. A same-named binding in another scope of the same file still reads as a use, which makes this under-report rather than over-report

• JSDoc/TSDoc extraction for any symbol

Related MCP server: ts-language-mcp

Tools

Structural Queries

Call Analysis

* Optional call_graph parameters:

• function - Focus on calls reachable from this function only

• direction - TD (top-down, default) or LR (left-right)

• include_external - Include calls to functions not defined in the file (default: false)

• scope - file (default) or package (cross-file analysis)

* Optional get_callers parameter:

• scope - file (default) or package (search all files in the directory)

Code Quality

A floating promise here means a discarded call to a function the same file declares async (or to fetch). Without type resolution that is the limit of what can be claimed soundly - an imported async function is not detected. It under-reports on purpose: the alternative is guessing from the callee's name, which reports map.get(k). | dead_code | Find unexported symbols never referenced inside their own file | path (directory), include_tests* | | find_implementations | Find classes satisfying an interface - explicit implements or type-checked structural match (semantic tier) | path, interface |

* Optional - omit to report all functions / exclude test files.

Documentation & Metadata

Multi-File Analysis

* Optional - include test files (default: false).

IDE Integration

Configuration

Claude Code

Add a .mcp.json file to the repository root:

{
  "mcpServers": {
    "ts-ast": {
      "command": "npx",
      "args": ["-y", "github:gclluch/ts-ast-mcp"]
    }
  }
}

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS):

{
  "mcpServers": {
    "ts-ast": {
      "command": "npx",
      "args": ["-y", "github:gclluch/ts-ast-mcp"]
    }
  }
}

GitHub Copilot (VS Code)

Create .vscode/mcp.json in the project root:

{
  "mcpServers": {
    "ts-ast": {
      "command": "npx",
      "args": ["-y", "github:gclluch/ts-ast-mcp"]
    }
  }
}

VSCode Extensions (Cline / Roo Code)

Add to the extension's MCP settings:

{
  "mcpServers": {
    "ts-ast": {
      "command": "npx",
      "args": ["-y", "github:gclluch/ts-ast-mcp"],
      "env": {}
    }
  }
}

Local development

For working on ts-ast-mcp itself, clone and build locally:

git clone https://github.com/gclluch/ts-ast-mcp.git
cd ts-ast-mcp
npm install    # prepare script builds automatically

Then point your .mcp.json at the local build:

{
  "mcpServers": {
    "ts-ast": {
      "command": "node",
      "args": ["/path/to/ts-ast-mcp/dist/index.js"]
    }
  }
}

Two-Tier Parsing

Most tools use the syntactic tier - ts.createSourceFile() parses a single file in under 5ms. No tsconfig or type checker needed.

Tools that operate across files (dead_code, analyze_package) and tools with scope: "package" (call_graph, get_callers) do this by re-parsing every file in scope on each call via the syntactic parser - always correct on the latest edit, at the cost of re-parsing repeatedly.

The semantic tier backs find_implementations, the one question that syntax cannot answer: whether a class satisfies an interface depends on member types and call signatures, not member names. loadProgram() (in src/parse.ts) builds a cached ts.Program and hands the tool the real type checker, which then answers via isTypeAssignableTo - the same rule the compiler applies to an implements clause.

Compiler options come from the nearest tsconfig.json; the file list never does. findConfigFile walks up, so the directory you asked about is routinely outside that config's include, and building from the config's own file list would produce a program that doesn't contain the files under analysis. The cache invalidates when the tsconfig's mtime changes or any file feeding the program changes, so an edit is always reflected on the next call while an unchanged tree reuses the Program.

Cost of the semantic tier is real: the first find_implementations call on a directory pays a full parse-bind-check pass. Subsequent calls on an unchanged tree are cache hits.

Arrow functions (const foo = () => {}) are detected as first-class functions throughout - they appear in list_functions, get_function_body, code_complexity, code_smells, and all other function-aware tools.

All output is plain text, not JSON.

Development

npm install
npm test        # builds, then runs the vitest suite

The suite covers the pure helpers (listTsFiles, bindingNames) plus an integration layer that spawns the real stdio server and drives it over JSON-RPC, so tool wiring is exercised the way a client actually uses it.

How to Verify

Once registered, you can ask your AI assistant to:

• "List all functions in Dashboard.tsx with their signatures."

• "Extract the full definition of the UserConfig interface."

• "Show me who calls the useApiQuery hook."

• "Generate a call graph for utils/ErrorUtils.ts."

• "What's the cyclomatic complexity of functions in DataTable.tsx?"

• "Which classes implement the DataProvider interface?"

• "Compare the old and new versions of api.ts structurally."

• "What AST node is at line 42, column 10?"

• "Give me a directory-level summary of src/hooks/."

• "Find error patterns in AuthService.ts."

• "Run a code smell check on BigComponent.tsx."

• "Find dead code in src/utils/."

• "What's the JSDoc for the useApiQuery function?"

• "List all exports from src/types/index.ts."