MCP 服务器

Graft

Local-first codebase context engine that parses code into a ranked dependency graph and serves it to AI tools via MCP for deep structural understanding.

README

<h1 align="center">Graft</h1>

Local-first codebase context engine for AI coding tools

Parse any codebase into a ranked dependency graph. Serve it to AI tools via MCP. Zero config. Zero cloud. Nothing leaves your machine.

Why Graft?

AI coding tools are only as good as the context they receive. Most tools either have no codebase understanding, or use proprietary indexing locked to a single editor.

Graft gives any MCP-compatible AI tool deep structural understanding of your codebase:

100K lines of code represented in ~2K tokens of ranked context
Dependency-aware — knows which files import what, and what depends on what
PageRank-scored — surfaces the most structurally important files first
Tool-agnostic — works with Claude, Cursor, Windsurf, Continue, or any MCP client
Local-first — no cloud, no telemetry, no code ever leaves your machine

$ npx graftmap map

src/indexer/pipeline.ts [score: 0.0842]
  function buildIndex (L19)

src/graph/pagerank.ts [score: 0.0731]
  function computePageRank (L52)
  function buildTeleportVector (L11)

src/parser/index.ts [score: 0.0654]
  function parseFiles (L28)

src/mcp/server.ts [score: 0.0612]
  function createGraftServer (L238)
  function startMcpServer (L327)

[~487 tokens]

Quick Start

# Run instantly with npx (no install needed)
npx graftmap map

# Or install globally
npm install -g graftmap
graft map

MCP Integration

Add Graft to your AI tool's MCP configuration. Once connected, your AI assistant automatically gets ranked codebase context.

Claude Desktop / Claude Code

Add to your MCP settings:

{
  "mcpServers": {
    "graft": {
      "command": "npx",
      "args": ["-y", "graftmap"],
      "cwd": "/path/to/your/project"
    }
  }
}

Cursor

Add to .cursor/mcp.json in your project:

{
  "mcpServers": {
    "graft": {
      "command": "npx",
      "args": ["-y", "graftmap"],
      "cwd": "/path/to/your/project"
    }
  }
}

Any MCP Client

Graft speaks standard MCP over stdio. Point any MCP-compatible client at npx graftmap and it works.

MCP Tools

Once connected, your AI tool gets these capabilities:

Tool	Description
`graft_map`	Ranked tree map of the codebase by structural importance
`graft_context`	Dependencies and definitions for a specific file
`graft_search`	Find definitions by name or kind (function, class, type, etc.)
`graft_impact`	Files affected by changing a given file
`graft_summary`	Project overview with key files and tech stack

Plus two MCP resources: graft://map and graft://file/{path}

CLI Commands

graft map                    # Ranked codebase tree
graft map --focus src/api.ts # Personalized view focused on a file
graft map --budget 4096      # Custom token budget
graft stats                  # File count, definitions, edges, cache age
graft impact src/auth.ts     # What breaks if you change this file?
graft search "handleRequest" # Find definitions by name
graft search "User" --kind class  # Filter by kind
graft serve                  # Start MCP server (default command)

How It Works

Your Codebase          Graft Pipeline              AI Tool
─────────────          ──────────────              ───────
  .ts .js .py    ──►   tree-sitter AST
                        extraction
                             │
                        dependency graph     ──►   graft_map
                        construction               graft_context
                             │                     graft_search
                        personalized         ──►   graft_impact
                        PageRank scoring           graft_summary
                             │
                        token-budgeted       ──►   Ranked context
                        rendering                  in ~2K tokens

Discover — finds all supported files, respects .gitignore
Parse — extracts definitions (functions, classes, types) and references using tree-sitter
Graph — builds a directed dependency graph (files as nodes, imports as edges)
Rank — runs personalized PageRank to score files by structural importance
Render — produces token-budgeted output that fits in any AI context window
Cache — stores results in .graft/cache.json for instant re-indexing

Supported Languages

Language	Definitions	References	Status
TypeScript	Functions, classes, interfaces, types, enums, methods	Imports, usages	Stable
JavaScript	Functions, classes, methods	Imports, usages	Stable
TSX/JSX	Same as TS/JS	Same as TS/JS	Stable
Python	Functions, classes, methods, decorators, dataclasses	Imports, usages	Stable
Go	—	—	Planned
Rust	—	—	Planned