Graph-Native
Multi-Agent AI
Memory lives in a property graph. Agents are graph walkers. Skills install from the internet at runtime. All written in Jac, the AI-native language built on Jaseci.
Get Started
One command. Works on Linux, macOS, and Windows.
curl -fsSL https://raw.githubusercontent.com/zero-abd/graphclaw/main/install.sh | bashEverything you need for AI agents
A complete platform where agents collaborate over shared graph memory, with skills that extend at runtime.
Graph Memory
Facts live as typed nodes with confidence scores that decay over time. The Dream walker prunes stale nodes, revalidates memories, and keeps topic tags fresh.
Multi-Agent Team
A Coordinator classifies intent and routes to DevOps, Planner, Builder, or Researcher. All agents share the same memory graph.
Skill Directory
Skills are self-contained modules with manifests. Install new skills at runtime from GitHub or the central registry — no restart needed.
Jac OSP
Nodes connected to root auto-persist. Per-user isolation is built in. Walkers traverse data instead of pulling it to logic.
5 Channels
Telegram, Discord, Slack, Email, and WhatsApp. One message bus routes everything through your agents.
Multi-Provider
OpenRouter, Anthropic, OpenAI, DeepSeek, Groq, Gemini, Mistral, Ollama, and any OpenAI-compatible endpoint.
Why Graphclaw?
Inspired by nanobot and openclaw — rebuilt from scratch with a graph-native core.
| nanobot | openclaw | Graphclaw | |
|---|---|---|---|
| Language | Python | TypeScript | Jac (compiles to Python) |
| Memory | Flat .md files | File-based | OSP property graph |
| Memory recall | LLM summarization | File scan | Indexed graph traversal + semantic |
| Agents | Single agent | Single agent | Multi-agent (5 specialists) |
| Skills | SKILL.md files | ClawHub | Dynamic directory + online install |
| Multi-user | No | No | Config scaffold present |
| Deployment | pip | pip | jac run → jac start → Kubernetes |
| AI functions | LLM calls | LLM calls | by llm() — Meaning Typed |
Multi-Agent Architecture
A Coordinator classifies your intent with by llm() and routes to the right specialist.
All agents share the same memory graph. A fact left by DevOps about a failed deployment is immediately visible to Coordinator on the next turn.
Memory is a Graph, Not a File
Every fact is a node with confidence that decays over time. The Dream walker maintains it automatically.
root
├──[:HasMemory]──▶ Memory { content, type, confidence: 0.9 }
├──[:HasSession]──▶ Session ──[:HasTurn]──▶ Turn
└──▶ Topic ──[:Tagged]──▶ Memory
Memory ──[:Relates]──▶ MemoryConfidence Decay
Every memory loses 0.01 confidence per day since last validation. After ~90 days without revalidation, a fact is considered stale and tombstoned.
Dual Recall
Substring match by default for speed. Semantic recall via by llm() on demand for deeper understanding.
Dream Walker
Runs every 2 hours: tombstones zero-confidence nodes, auto-tags untagged memories, revalidates still-accurate decaying memories.
Extensible Skill System
Two skill types that extend your agents at runtime — no restart needed.
Native Skills
Fast, typed Python modules with a JSON manifest.
ClawHub Skills
13,000+ community skills from the ClawHub registry.
One Bus, 5 Channels
Connect your agents to any platform. One message bus routes everything.