graqle
Dev intelligence layer that builds a knowledge graph from any codebase and exposes 7 MCP tools for graph-powered reasoning, impact analysis, and preflight safety and governance checks.
README
<!-- mcp-name: com.graqle/graqle --> <div align="center">
graQle
Query your architecture, not your files.
The dev intelligence layer that turns any codebase into a self-learning knowledge graph.<br/> 2,000+ tests. 396 modules compiled. 201 skills. Zero cloud required.
Website · PyPI · GitHub · Changelog
</div>
3 commands. That's it.
pip install graqle[api]
graq init
graq reason "what breaks if I change auth?"
3 nodes activated. 500 tokens. 5 seconds. $0.0003.
Not 60 files. Not 50,000 tokens. Not $0.15. Not guessing.
What Graqle does
Your AI assistant reads files one at a time. It doesn't understand your architecture — it reads text.
Graqle builds a knowledge graph of your entire codebase. Every module becomes a reasoning agent. When you ask a question, only the relevant nodes activate, reason about their domain, and synthesize one answer.
pip install graqle → graq init → Knowledge Graph
│
┌──────────┬──────────────┼──────────┐
▼ ▼ ▼ ▼
CLI Python SDK MCP Server Studio UI
(terminal) (scripts) (any IDE) (dashboard)
The graph is the product. Your machine, your API keys, your data.
What's new in v0.26.0
Intelligence Compilation
Graqle now compiles your knowledge graph into actionable intelligence:
graq compile # Compile intelligence from your graph
This produces a .graqle/intelligence/ directory with:
- 396 module packets — risk scores, impact radius, consumers, dependencies
- 135 insights — warnings, suggestions, connections, superlatives
- Risk heatmap data — LOW/MEDIUM/HIGH/CRITICAL per module
- CLAUDE.md auto-injection — your AI assistant learns your architecture automatically
Governance Gate (DRACE)
Every reasoning session is scored on 5 axes: Data quality, Relevance, Accuracy, Completeness, Evidence strength.
graq verify # Pre-commit governance check
- Hash-chained audit trails (tamper-evident)
- Evidence chains linking decisions to source
- Scope gates preventing out-of-domain reasoning
- Auto-recompile when intelligence goes stale
Multi-Signal Activation (Neo4j)
Gate + Rerank: semantic score gates everything, topology signals only amplify.
final = semantic × (1 + authority + memory + link + freshness)
Max amplification: 1.45×. An irrelevant node can never sneak through.
Self-Learning Loop
graq learn → graph grows → git commit → auto-recompile → CLAUDE.md updates → AI gets smarter
The graph remembers which nodes produce useful answers. Gets better with every query.
Studio Dashboards
6 new visual dashboards: Intelligence, Governance, Health, Learning, Control, Share.
Works with everything
| IDE / Tool | How | Setup |
|---|---|---|
| Claude Code | MCP server + CLAUDE.md | graq init (auto) |
| Cursor | MCP server + .cursorrules | graq init --ide cursor |
| VS Code + Copilot | MCP server | graq init --ide vscode |
| Windsurf | MCP server + .windsurfrules | graq init --ide windsurf |
| JetBrains / Codex / Replit | CLI + SDK | graq init --ide generic |
| CI/CD pipelines | Python SDK | pip install graqle |
No cloud account. No infrastructure. No config files to write.
CLI
# Reasoning
graq reason "what depends on auth?" # Graph reasoning
graq context auth-lambda # 500-token focused context
graq inspect --stats # Graph statistics
graq "what is safe to refactor?" # Natural language (auto-routed)
# Build & compile
graq init # Scan repo, build graph, wire IDE
graq scan repo . # Rescan codebase
graq scan all . # Code + JSON + documents
graq compile # Compile intelligence layer
graq verify # Governance gate + staleness check
# Teach
graq learn node "auth-service" --type SERVICE
graq learn edge "Payments" "auth" -r DEPENDS_ON
graq learn discover --from "auth-service" # Auto-discover connections
graq learn doc architecture.pdf # Document ingestion
# Studio & server
graq studio # Launch visual dashboard
graq serve # Start REST API
graq doctor # Health check
Python SDK
from graqle.core.graph import Graqle
from graqle.backends.api import AnthropicBackend
graph = Graqle.from_json("graqle.json")
graph.set_default_backend(
AnthropicBackend(model="claude-sonnet-4-6")
)
result = graph.reason(
"What services depend on auth?",
max_rounds=3,
strategy="top_k"
)
print(result.answer)
print(f"Confidence: {result.confidence:.0%}")
print(f"Cost: ${result.cost_usd:.4f}")
MCP Tools
Available automatically in Claude Code, Cursor, VS Code, and Windsurf after graq init:
| Tool | What it does |
|---|---|
graq_context |
500-token focused context (replaces reading entire files) |
graq_reason |
Multi-agent graph reasoning |
graq_impact |
"What breaks if I change X?" |
graq_preflight |
Pre-change safety check |
graq_lessons |
Surface past mistakes before you repeat them |
graq_learn |
Teach the graph new knowledge |
graq_inspect |
Graph structure inspection |
graq_reload |
Hot-reload graph without restarting |
Studio Dashboard
graq studio # Opens http://127.0.0.1:8888/studio/
| Page | What you see |
|---|---|
| Intelligence | Risk heatmap, module packets, 135 insights, impact matrix |
| Governance | DRACE radar chart, audit timeline, evidence chains, shareable badge |
| Health | Streak calendar, improvement suggestions, graph health trends |
| Learning | Skill activations, domain breakdown, recompile history |
| Graph Explorer | Interactive force-directed graph with intelligence overlay |
| Control | Multi-instance management, cross-repo insights |
14 backends, one config line
# graqle.yaml
model:
backend: groq # or: anthropic, openai, bedrock, gemini, ollama, deepseek, mistral, ...
| Backend | Cost | Setup |
|---|---|---|
| Ollama | $0 (local) | backend: ollama |
| Anthropic | ~$0.001/query | backend: anthropic |
| OpenAI | ~$0.001/query | backend: openai |
| AWS Bedrock | AWS pricing | backend: bedrock |
| Google Gemini | ~$0.0001/query | backend: gemini |
| Groq | ~$0.0005/query | backend: groq |
| DeepSeek | ~$0.0001/query | backend: deepseek |
| Mistral | ~$0.0002/query | backend: mistral |
| Together | ~$0.0005/query | backend: together |
| OpenRouter | Varies | backend: openrouter |
| Fireworks | ~$0.0005/query | backend: fireworks |
| Cohere | ~$0.0003/query | backend: cohere |
| vLLM | Your GPU | backend: local |
| llama.cpp | $0 (CPU) | backend: local |
Task-based routing
Different models for different tasks. Fast models for lookups, smart models for reasoning:
routing:
default_provider: groq
rules:
- task: reason
provider: anthropic
model: claude-sonnet-4-6
- task: context
provider: groq
model: llama-3.1-8b-instant
Auto-scaling backend
Graqle starts with JSON/NetworkX (zero deps). When your graph grows:
| Graph size | Backend | What happens |
|---|---|---|
| < 5,000 nodes | JSON/NetworkX | Default. Instant. Zero config. |
| 5,000+ nodes | Neo4j | Auto-recommended. Migration handled. |
| Team/Enterprise | Neo4j + GDS | Vector search, PageRank, Adamic-Adar proximity. |
pip install graqle[neo4j] # Adds neo4j driver — same API, just faster
Document-aware intelligence
Graqle connects code to documents to configs in one graph.
| Source | Formats | What it extracts |
|---|---|---|
| Code | Python, TypeScript, JavaScript, Go, Rust, Java | Functions, classes, modules, imports, call graphs |
| Documents | PDF, DOCX, PPTX, XLSX, Markdown, TXT | Sections, decisions, requirements, stakeholders |
| JSON | package.json, OpenAPI, tsconfig, CDK, SAM | Dependencies, endpoints, infrastructure resources |
Auto-linking: exact match, fuzzy match, semantic match, LLM-assisted. Cross-source deduplication. Contradiction detection.
The numbers
| Metric | Graqle | Reading files |
|---|---|---|
| Tokens per query | 500 | 50,000 |
| Cost per query | $0.0003 | $0.15 |
| Time to answer | <5 seconds | 20 minutes |
| Tests passing | 2,000+ | — |
| Modules compiled | 396 | — |
| Skills available | 201 | — |
Installation
pip install graqle # Minimal (no API keys needed)
pip install graqle[api] # + Anthropic, OpenAI, Bedrock
pip install graqle[docs] # + PDF, DOCX, PPTX, XLSX
pip install graqle[neo4j] # + Neo4j graph database
pip install graqle[all] # Everything
| Extra | What it adds |
|---|---|
api |
anthropic, openai, boto3 |
docs |
pdfplumber, python-docx, python-pptx, openpyxl |
neo4j |
neo4j driver |
embeddings |
sentence-transformers |
gpu |
torch, transformers, peft, vllm |
cpu |
llama-cpp-python |
studio |
fastapi, uvicorn, jinja2 |
dev |
pytest, ruff, mypy, coverage |
15 Innovations (Patent EP26162901.8)
Every innovation is free under Apache 2.0.
| # | Innovation | Why it matters |
|---|---|---|
| 1 | Chunk-level semantic scoring | Finds the exact function, not the file |
| 2 | Zero-cost transparency | See which agents activated and why |
| 3 | Convergent message passing | Agents discuss until they agree |
| 4 | Backend fallback chain | Auto-switches models on failure |
| 5 | Topology-aware synthesis | Answers reflect graph structure |
| 6 | 3-layer governance (DRACE) | Domain boundaries on every output |
| 7 | Formula-based compliance | Quantitative scores, not pass/fail |
| 8 | Auto-generate ontologies | OWL+SHACL from your codebase |
| 9 | Adaptive activation | 3 nodes for simple, 50 for complex |
| 10 | Cross-query learning | Remembers what works for which patterns |
| 11 | Per-entity model selection | Security nodes use capable models |
| 12 | Retrieval-to-reasoning pipeline | Documents → graph reasoning |
| 13 | Hybrid skill matching | Regex precision + semantic flexibility |
| 14 | Neo4j vector + graph search | Embedding similarity + graph traversal in one query |
| 15 | Activation memory | Persistent cross-session node effectiveness |
Quick start
# Python project
pip install graqle[api] && cd my-project && graq init
graq reason "what depends on the database module?"
# With documents
pip install graqle[api,docs]
graq scan all . && graq scan wait
graq reason "what does the architecture doc say about auth?"
# Claude Code (auto-wired)
graq init # Claude Code now has graq_context, graq_reason, graq_impact, etc.
# Cursor
graq init --ide cursor
Pricing
Free for individuals. Always.
| Free ($0) | Pro ($19/mo) | Team ($29/dev/mo) | |
|---|---|---|---|
| All 15 innovations | Yes | Yes | Yes |
| CLI + SDK + MCP + API | Yes | Yes | Yes |
| 14 LLM backends | Yes | Yes | Yes |
| Document scanning | Yes | Yes | Yes |
| Intelligence compilation | Yes | Yes | Yes |
| Studio dashboards | Basic | Full | Full + team |
| Governance (DRACE) | Current session | 30-session history | Unlimited |
| Audit trails | 3 sessions | 20 sessions | Unlimited |
| Health streaks | 7 days | Full year | Team streaks |
| Commercial use | Yes | Yes | Yes |
Citation
@article{kumar2026graqle,
title = {Graqle: Governed Intelligence through Graph-of-Agents Reasoning
over Knowledge Graph Topologies with Semantic SHACL Validation},
author = {Kumar, Harish},
year = {2026},
institution = {Quantamix Solutions B.V.},
note = {European Patent Application EP26162901.8},
url = {https://github.com/quantamixsol/graqle}
}
Contributing
git clone https://github.com/quantamixsol/graqle
cd graqle
pip install -e ".[dev]"
pytest # 2,000+ tests
See CONTRIBUTING.md for development setup, testing, and PR guidelines.
License
Apache 2.0 — use it commercially, modify it freely, keep the attribution.
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。