MindGraph
Enables AI agents to query and analyze code across multiple repositories through a unified knowledge graph, with tools for symbol search, impact analysis, and graph algorithms.
README
MindGraph
Multi-Repository Code Knowledge Graph Engine
MindGraph is a local-first static analysis engine that turns N source repositories into a single, queryable code knowledge graph and exposes it to AI agents through the Model Context Protocol (MCP). It unifies the core capabilities of CodeGraph (symbol graph + impact analysis), GitNexus (trace, route mapping, change detection, cross-repo groups), and Graphify (community detection, god nodes, shortest path, edge confidence) into one tool with first-class multi-repo support.
Features
- Multi-repo merging — Index an unbounded number of repositories into one unified graph. Each repo retains its identity (
repo_id,repo_name,root_path) while participating in a shared symbol space. - 22 MCP tools spanning exploration, advanced graph queries, graph intelligence, and group management (see MCP Tools Reference).
- Cross-repo dependency resolution —
#includes, calls, and type references that span repositories are resolved into explicit cross-repo edges usingcross_repo_includesmappings. - Layer architecture annotations — Tag repos with a
layer(e.g.ui,business,engine,sdk) and monorepo directories withsub_layers, then filter or group any query by layer. - Edge confidence tagging — Every edge carries a
confidenceofEXTRACTED(from source),INFERRED(heuristic), orAMBIGUOUS, plus aprovenance(static,inferred,cross-repo,synthesized). - Graph intelligence — Leiden/Louvain community detection, god-node ranking (degree / betweenness / PageRank), and shortest-path queries (directed and undirected).
- C/C++ focused, multi-language aware — C and C++ are the primary test targets; TypeScript, Python, Rust, Go, and Java extractors are included.
- Local and private — All parsing, storage, and querying happens on your machine. No network calls during indexing or querying.
- Incremental re-index — Changing one repo only re-indexes that repo and re-resolves its cross-repo edges. Circular repo dependencies are detected and reported, never infinite-looped.
Quick Start
Installation
Requirements: Node.js >= 22, npm.
git clone https://github.com/your-org/mindgraph.git
cd mindgraph
npm install
npm run build
npm link # optional: exposes the `mindgraph` binary on PATH
Index a single repository
mindgraph index /path/to/repo --layer engine --id engine-core
Index a multi-repo group
Create a group_config.json (see Group Configuration), then:
mindgraph index ./group_config.json
MindGraph indexes every repo in the group and resolves cross-repo edges in one pass.
Start the MCP server
mindgraph mcp --group-config ./group_config.json
The server speaks MCP over stdio and is intended to be launched by an MCP client (Claude Desktop, Cursor, etc.), not run interactively.
Configure your MCP client
See MCP Configuration below.
MCP Configuration
Add MindGraph to your MCP client's config file.
Claude Desktop — ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):
{
"mcpServers": {
"mindgraph": {
"command": "mindgraph",
"args": ["mcp", "--group-config", "/absolute/path/to/group_config.json"],
"env": {
"MINDGRAPH_DIR": "/absolute/path/to/data-dir"
}
}
}
}
Cursor — .cursor/mcp.json in your workspace (or global Cursor MCP settings):
{
"mcpServers": {
"mindgraph": {
"command": "mindgraph",
"args": ["mcp", "--group-config", "/absolute/path/to/group_config.json"]
}
}
}
If MindGraph isn't on PATH, use the absolute path to the binary or invoke it through Node:
{
"mcpServers": {
"mindgraph": {
"command": "node",
"args": ["/absolute/path/to/mindgraph/dist/index.js", "mcp"]
}
}
}
Data directory resolution order: MINDGRAPH_DIR env var, then ./.mindgraph (if present), then ~/.mindgraph.
Group Configuration
A group_config.json declares the repos in a group, their architectural layers, and the include-path mappings used to resolve cross-repo references.
{
"name": "av-editor",
"repos": [
{
"id": "ui-layer",
"path": "/path/to/ui-repo",
"layer": "ui",
"include_paths": ["src/", "include/"]
},
{
"id": "engine-core",
"path": "/path/to/engine-core",
"layer": "engine",
"sub_layers": ["logic", "render", "service", "base", "gpu"],
"include_paths": ["include/", "src/"]
}
],
"cross_repo_includes": [
{
"from": "ui-layer",
"to": "engine-core",
"mapping": { "engine/": "/path/to/engine-core/include/" }
}
]
}
| Field | Description |
|---|---|
name |
Group identifier. |
repos[].id |
Stable repo identifier used in queries and edges. |
repos[].path |
Absolute path to the repository root. |
repos[].layer |
Architectural layer tag (e.g. ui, engine, sdk). |
repos[].sub_layers |
Optional sub-layer names for monorepo-style directories. |
repos[].include_paths |
Directories searched when resolving #includes. |
cross_repo_includes[] |
Maps an include prefix in from to an absolute path inside to. |
Generate a starter config with:
mindgraph group init --name av-editor --repo /path/to/ui --repo /path/to/engine
CLI Reference
Global options: --db <dir> (override data directory), -v, --verbose (debug logging).
mindgraph index [path]
Index a repository directory or a group_config.json.
# Single repo
mindgraph index /path/to/repo --layer engine --id engine-core --group default
# Multi-repo group
mindgraph index ./group_config.json
Options: --layer <layer>, --id <id>, --group <name> (default: default).
mindgraph mcp
Start the MCP server over stdio.
mindgraph mcp --group-config ./group_config.json
mindgraph status
Show per-repo and aggregate index statistics.
mindgraph status
mindgraph status --repo engine-core
mindgraph search <query>
Quick symbol search from the terminal.
mindgraph search "VideoEncoder" --kind class --limit 20 --repo engine-core
mindgraph group
Manage repository groups.
# Create a group_config.json in the current directory
mindgraph group init --name my-group --repo /path/to/a --repo /path/to/b
# Add (and index) a repo into a group
mindgraph group add /path/to/repo --group my-group --layer sdk --id sdk-repo
# Re-sync cross-repo edges (incremental, or --full to re-index first)
mindgraph group sync --group my-group
mindgraph group sync --group my-group --full --group-config ./group_config.json
MCP Tools Reference
All tools are read-only, idempotent, and non-destructive (except mindgraph_rename, which is dry-run by default).
Core (CodeGraph)
| Tool | Description |
|---|---|
mindgraph_explore |
Primary semantic exploration; returns symbols, call paths, code excerpts, and cross-repo relationships for a natural-language or symbol query. |
mindgraph_search |
Quick symbol search by name/pattern, filterable by kind and repo. |
mindgraph_node |
Detailed info for one symbol: signature, docstring, visibility, callers/callees counts, optional source code. |
mindgraph_callers |
Who calls this symbol, with call sites and configurable depth. |
mindgraph_callees |
What this symbol calls, with call sites and configurable depth. |
mindgraph_impact |
Blast-radius analysis: depth-grouped impact tree with confidence scores and affected repos/layers. |
mindgraph_files |
Indexed file tree with language, symbol counts, and layer annotations. |
mindgraph_status |
Index health: per-repo and aggregate files, nodes, edges, unresolved refs, last sync. |
Advanced (GitNexus)
| Tool | Description |
|---|---|
mindgraph_trace |
Shortest directed path between two symbols over selected edge types. |
mindgraph_detect_changes |
Git-diff impact analysis: changed files → affected symbols → downstream impact, grouped by layer. |
mindgraph_rename |
Coordinated multi-file rename, cross-repo aware, dry-run by default. |
mindgraph_route_map |
API/framework route map with handler symbols and middleware chains. |
mindgraph_cypher |
Raw Cypher-style query over the underlying SQLite graph. |
Graph Intelligence (Graphify)
| Tool | Description |
|---|---|
mindgraph_communities |
Louvain/Leiden community detection with cross-community bridges and layer distribution. |
mindgraph_god_nodes |
Most-connected core abstractions, ranked by degree, betweenness, or PageRank. |
mindgraph_shortest_path |
Conceptual undirected shortest path with edge-type and confidence annotations. |
mindgraph_graph_stats |
Global statistics: node/edge counts by kind, communities, confidence breakdown, layer distribution, cross-repo edge count. |
mindgraph_neighbors |
Direct neighbors of a symbol, grouped by relation type with edge metadata. |
Multi-Repo Management
| Tool | Description |
|---|---|
mindgraph_group_list |
List configured groups with their repos, layers, and index status. |
mindgraph_group_sync |
Rebuild cross-repo links for a group (incremental or full). |
mindgraph_repo_add |
Add and index a repository in a group. |
mindgraph_repo_remove |
Remove a repository from a group and clean up its data. |
Architecture
| Layer | Technology |
|---|---|
| Runtime | Node.js + TypeScript (ESM) |
| Parser | web-tree-sitter with WASM grammars (no native build required) |
| Storage | SQLite (mindgraph.db) with FTS5 full-text search |
| Graph algorithms | graphology + communities-louvain, metrics, shortest-path |
| MCP server | @modelcontextprotocol/sdk over stdio |
| CLI | commander |
| Schema validation | zod |
| Tests | vitest |
Pipeline: scan (file discovery + ignore rules) → parse (tree-sitter) → extract (per-language symbol/edge extractors) → resolve (intra-repo references) → cross-repo resolve (group-aware edge synthesis) → store (SQLite + FTS5) → serve (MCP tools).
Source layout:
src/
├── index.ts # CLI entry
├── mcp/ # MCP server + one file per tool
├── core/ # scanner, parser, language extractors, resolvers
├── graph/ # SQLite store, traversal, communities, centrality, paths
├── group/ # group config, multi-repo orchestration, sync
└── utils/ # logger, config, fs helpers
See SPEC.md for the full data model (SQLite schema) and the cross-repo resolution algorithm.
Testing
npm test # run the full vitest suite once
npm run test:watch # watch mode
npm run lint # eslint over src/
Integration tests run against a synthetic 3-repo C++ fixture under test/fixtures/; end-to-end tests target real C++ projects (see SPEC.md §5).
License
MIT — see LICENSE for details.
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。