Add it to Claude Code
claude mcp add graqle -- graq mcpMake your agent remember this setup
graqle's config, env vars, and the gotchas you hit — recalled in every future Claude Code, Cursor, and Codex session.
npx conare@latestFree · one command · indexes the sessions already on disk. Set up in the browser instead →
What it does
- Builds a persistent knowledge graph from any codebase
- Graph-of-agents reasoning for architecture-aware answers
- Automated impact analysis and blast radius calculation
- DRACE governance scores for preflight safety checks
- Supports 14 LLM backends including local Ollama
Tools 5
graq_scanScans the current repository to build or update the knowledge graph.graq_runExecutes a natural language query against the codebase knowledge graph.graq_impactPerforms impact analysis to determine the blast radius of a file change.graq_preflightRuns governance and safety checks on proposed refactors.graq_learnAdds new institutional knowledge to the persistent graph.Environment Variables
GRAQ_API_KEYAPI key for GraQle services if required by the backend configuration.Try it
Original README from quantamixsol/graqle
Your AI reads files. Gra**Q**le reads architecture.
The context layer for AI coding agents. Scan any codebase into a knowledge graph. Every module becomes an agent. Ask questions — get architecture-aware answers in 5 seconds, not 2 minutes.
pip install graqle && graq scan repo . && graq run "what's the riskiest file to change?"
Website · Dashboard · PyPI · Changelog
50,000 tokens → 500 tokens. Same answer.
| Without GraQle | With GraQle | |
|---|---|---|
| "What depends on auth?" | AI reads 60 files, guesses | Graph traversal → exact answer in 5s |
| Tokens per question | 50,000 | 500 |
| Cost per question | ~$0.15 | ~$0.0003 |
| Impact analysis | Manual grep + hope | graq impact auth.py → full blast radius |
| Memory across sessions | Lost when chat resets | Persistent knowledge graph |
| Confidence in answers | "I think..." | Confidence score + evidence chain |
"We scanned 17,418 nodes across 3 projects in one session. Found 807 jargon blind spots, 218 ghost UI elements, and a CTA that was 20px tall (44px minimum). Cost: $0.30." — Quantamix Website Audit
How it works — 60 seconds
# 1. Install
pip install graqle
# 2. Scan your codebase into a knowledge graph
graq scan repo .
# → 2,847 nodes, 9,156 edges — your entire architecture mapped
# 3. Ask anything about your architecture
graq run "explain the payment flow end to end"
# → Graph-of-agents activates 8 relevant nodes, synthesizes answer
# → Confidence: 92% | Cost: $0.001 | Time: 5.2s
# 4. Connect to your AI IDE (zero config change)
graq init # Claude Code, Cursor, VS Code, Windsurf — auto-detected
Your AI now has 27 architecture-aware MCP tools — including Phantom computer skills for live browser automation. No workflow change — it uses them automatically.
What makes Graqle different
<table> <tr> <td width="50%">🔬 Graph-of-Agents Reasoning
Every module in your codebase becomes an autonomous agent. When you ask a question, only the relevant agents activate — they debate, exchange evidence, and synthesize one answer with a confidence score and full audit trail.
This is not RAG. This is structured multi-agent reasoning over your dependency graph.
</td> <td width="50%">🧠 The Graph Learns
graq learn "auth requires refresh token rotation"
graq grow # Auto-runs on git commit
Every interaction makes the graph smarter. Lessons persist across sessions. New developers and AI tools inherit your team's institutional knowledge automatically.
</td> </tr> <tr> <td>🛡️ Governed AI Decisions
graq preflight "refactor the database layer"
# → 4 modules depend on connection pool
# → 2 have no tests
# → DRACE score: 0.72 (proceed with caution)
Every answer is auditable. DRACE governance scores across 5 dimensions. Full evidence chains. Patent-protected.
</td> <td>⚡ 14 LLM Backends
model:
backend: ollama # Free, offline, air-gapped
# Also: anthropic, openai, groq, deepseek,
# gemini, bedrock, together, mistral,
# fireworks, cohere, openrouter, vllm, custom
Use your own API keys. Run fully offline with Ollama. Smart routing assigns different models to different tasks.
</td> </tr> </table>Real stories from production
<details> <summary><b>📊 "807 jargon blind spots in 90 seconds"</b> — Website audit with SCORCH</summary>A professional website with WCAG AAA compliance still had 807 unexplained acronyms (TAMR+, TRACE, SHACL, HashGNN) that compliance officers would bounce on. GraQle's SCORCH engine found them all in one scan. Lighthouse missed every one.
Before: "Explore our TAMR+ SHACL-compliant governance pip