intelligraph-mini
# Intelligraph-mini
Local-first MCP graph intelligence server — same RRF hybrid search, multi-hop traversal, source snippets, and rationale nodes as the full Intelligraph platform, but without Docker, web UI, SSO, or chat. Just tools for your AI agent.
> **Need the full platform?** [Intelligraph](https://github.com/kfireew/Intelligraph) adds Docker, React web UI, chat completions, SSO/PKCE, closed-network deployment, and tuning controls.
## Quick start
```bash
pip install intelligraph-mini
# In your project directory:
intelligraph-mini --repo-dir .
# Or with MCP config (.mcp.json):
{
"mcpServers": {
"intelligraph-mini": {
"command": "intelligraph-mini",
"args": ["--repo-dir", "."]
}
}
}
```
First run builds graphify + CRG indexes (~60s). Subsequent runs load cached (~2s). The bundled `all-MiniLM-L6-v2` model (87MB) works fully offline — no API calls, no network.
## Tools
| Tool | Description |
|------|-------------|
| `search(query)` | RRF hybrid search (FTS5 + semantic embeddings). Finds symbols by meaning. |
| `node(name, depth=2)` | Multi-hop subgraph + source code snippets + rationale notes. |
| `path(from, to)` | Shortest path between two symbols in the call graph. |
| `impact(name)` | Blast-radius analysis over CALLS/IMPORTS_FROM edges. |
| `local_files(paths)` | Read source files from disk. |
## How it works
1. **Build** (first run): `graphify update .` + `code-review-graph build` → `graphify-out/graph.json` + `.code-review-graph/graph.db`
2. **Snippets**: reads source files, stores ~500 char snippets per node in `node_snippets` table
3. **Search**: RRF (Reciprocal Rank Fusion, k=30) blends FTS5 keyword ranking with embedding cosine similarity. Adaptive 50% cutoff returns only genuinely relevant files.
4. **Traversal**: BFS with token budget over cached adjacency (scales to 140k edges)
5. **Rationale**: surfaces `#NOTE`/`#WHY` nodes from graphify's rationale extraction
## Requirements
- Python 3.10+
- `graphifyy` and `code-review-graph` CLIs on PATH (installed automatically as dependencies)
## License
MIT
TDQS
Scored across 5 tools
Each tool has a distinct role: search locates symbols, node shows connections, path finds shortest routes, impact assesses blast radius, and local_files reads source. No two tools overlap in purpose.
Names are mostly single-word and lowercase, but mix verbs (search) and nouns (node, path, impact, local_files). The underscore in local_files is inconsistent with the other names, though still readable.
Five tools is well-scoped for a mini code-graph intelligence server, covering search, exploration, pathfinding, impact analysis, and file reading without redundancy.
The set covers the full analysis workflow: locate, explore, trace, assess, and read. Instructions to use built-in Read with line ranges fill any need for source access, leaving no obvious gaps.