GraphClaw
Work is a graph
A property-graph model captures dependencies, blockers, owners, and goals — so the agent can reason about critical paths and bottlenecks.
GraphClaw is a graph-based task orchestration system where an AI agent manages work for humans and other agents — with explicit dependencies, scored priorities, and an auditable lifecycle. Self-hosted and pluggable to the core.
Real work has dependencies, owners, delegation chains, and goal context. Modeling it as a property graph lets the agent prioritize, delegate, and explain — instead of just tracking.
A property-graph model captures dependencies, blockers, owners, and goals — so the agent can reason about critical paths and bottlenecks.
Priorities come from timeline urgency, dependency weight, critical path, blockers, human overrides, and resource risk — and every score is explainable.
The agent breaks down work into subgraphs, delegates to humans and AI, follows up, batches outreach, and surfaces prioritized briefings.
Reachable from Web chat, Email, WhatsApp, Telegram, Slack, and Teams. The agent is fully useful without ever opening a visual interface.
SSE event streams and WebSocket chat keep every surface current. Whatever changes — in chat, cockpit, or API — reflects in the graph instantly.
Database, gateway, LLM-provider, and infrastructure layers sit behind interfaces — swap Postgres+AGE, providers, channels, or storage without rewriting core logic.
The conversational agent is the primary interface. The visual cockpit and the admin panel are power-user complements — never gatekeepers.
Daily briefings, quick decisions, task creation, and status updates over chat or email. Designed so you never need to open a UI.
Primary · dailyReview and edit graph structure, run planning sessions, decompose projects, and visualize dependencies. Add, edit, or override anything — it's your data.
Power use · weeklyConfigure channels, skills, MCP registry, LLM providers, scoring weights, guardrails, and SSO. The policy layer that everything else operates within.
OccasionalPrioritization is transparent, and every task moves through an auditable lifecycle.
Weighted contributions for a single task — surfaced, not hidden behind an advanced mode.
Tasks move through explicit states with recorded transitions and cascade logic.
Override scoring, lock nodes from agent modification, or edit any node and edge directly. The agent assists — it never gatekeeps.
Every external dependency sits behind an interface, so you can run GraphClaw your way.
GraphClaw runs as a self-hosted stack you control. Start with the backend and cockpit guides, or follow the public roadmap to see what's Now / Next / Later.