graphclaw 0.2.3

Open-source graph-based task orchestration with AI agents

GraphClaw — Graph-Based Task Orchestration System

Domain: graphclaw.ai | GitHub: graphclaw/graphclaw

GraphClaw is an open-source graph-based task orchestration system where an AI agent manages tasks for humans and other agents via a property graph. Tasks, goals, constraints, and resources are modeled as graph nodes connected by typed edges (dependencies, assignments, blocking relationships). A 7-factor scoring algorithm continuously prioritizes the action queue, while a state machine enforces lifecycle invariants.

Architecture

┌──────────────────────────────────────────────────────────────┐
│          CLI (Typer + Rich) / Gateway API (FastAPI)          │
│   task list/show/create │ agent run/score/briefing │ /docs   │
├──────────────────────────────────────────────────────────────┤
│                  Channel Gateway Layer                       │
│  ChannelAdapter ABC │ Email (done) │ WhatsApp/Telegram (P2)  │
├────────────────────────┬─────────────────────────────────────┤
│   Agent Reasoning Loop │           Skill Runtime             │
│   fetch → score → act  │  SkillWorker │ SKILL.md │ LLMClient │
├────────────────────────┴─────────────────────────────────────┤
│                   LLM Provider Layer                         │
│  LLMClient ABC │ Anthropic │ OpenAI │ LiteLLM (pluggable)    │
├──────────────────────────────────────────────────────────────┤
│   State Machine │ Scoring Engine (7 factors) │ Trigger Engine│
├──────────────────────────────────────────────────────────────┤
│          Domain Models (Pydantic v2) — 17 node types         │
├──────────────────────────────────────────────────────────────┤
│   Database Layer — GraphStore ABC                            │
│   Postgres + Apache AGE (graph) + pgvector (embeddings)      │
└──────────────────────────────────────────────────────────────┘

Plugin Architecture

GraphClaw is designed as a pluggable 4-layer system. Every infrastructure concern uses the ABC + Factory + Strategy pattern so backends are swappable without changing business logic:

Layer	ABC	How to add a backend
Database	GraphStore + GraphQueryEngine	Implement ABC, place in src/graphclaw/db/<name>/
Gateway	ChannelAdapter	Implement ABC, place in src/graphclaw/gateway/channels/<name>/
LLM	LLMClient	Implement ABC, place in src/graphclaw/llm/<name>/
Infra	StorageClient, MessageBroker, SecretsClient	Implement ABC, register in factory

See docs/architecture.md for the full design.

Tech Stack

Component	Technology
Language	Python 3.12+
Graph DB	PostgreSQL + Apache AGE (Cypher queries)
Vectors	pgvector (embedding similarity)
API	FastAPI + Swagger UI (/docs)
AI Orchestration	Anthropic SDK / OpenAI SDK (pluggable via LLMClient ABC)
Multi-provider	LiteLLM (default), Anthropic, OpenAI (all swappable)
Validation	Pydantic v2
CLI	Typer + Rich
Infrastructure	Docker Compose
Storage	MinIO (S3-compatible, local dev)
Caching/Broker	Redis (local dev)

Quick Start

Prerequisites

• Docker & Docker Compose
• Python 3.12+

1. Clone and start the stack

git clone https://github.com/graphclaw/graphclaw
cd graphclaw
cp docker/.env.example docker/.env
# Edit docker/.env and fill in real values (LLM keys, OAuth credentials, etc.)
docker compose -f docker/docker-compose.yml up -d

Note: this Docker workflow reads environment variables from docker/.env. For non-Docker local runs (for example CLI or direct Python execution), use the repository root .env.

This starts Postgres+AGE+pgvector, MinIO, Redis, and the gateway service.

2. Install the project

pip install -e ".[dev]"

3. Run tests

# Unit tests (no DB required) — 485+ tests
pytest tests/ --ignore=tests/test_db -q

# Integration tests (requires running DB)
# Replace <your-db-password> with the value of DB_PASSWORD from docker/.env
export TEST_DATABASE_URL=postgresql://graphclaw:<your-db-password>@localhost:5432/graphclaw_test
pytest tests/test_db/ -m integration

4. Use the CLI

# Score tasks and show action queue
graphclaw agent score

# Generate a briefing
graphclaw agent briefing

# Run one agent reasoning cycle
graphclaw agent run

5. Gateway API (Swagger UI)

Start the gateway service, then visit: http://localhost:8080/docs

# Or run directly
python -m graphclaw.gateway.app

Project Structure

graphclaw/
├── src/graphclaw/
│   ├── agent/              # Orchestrating agent reasoning loop
│   ├── cli/                # CLI interface (Typer + Rich)
│   ├── db/                 # Database layer (GraphStore ABC + AGE backend)
│   │   ├── base.py         # GraphStore + GraphQueryEngine ABCs
│   │   ├── factory.py      # create_graph_store()
│   │   └── age/            # Postgres+AGE backend
│   ├── gateway/            # Channel gateway (ChannelAdapter ABC)
│   │   ├── channel_base.py # ChannelAdapter ABC
│   │   ├── channel_registry.py
│   │   └── channels/email/ # Email plugin (IMAP+SMTP)
│   ├── llm/                # LLM provider abstraction (LLMClient ABC)
│   │   ├── base.py         # LLMClient ABC + LLMMessage/LLMResponse types
│   │   ├── factory.py      # create_llm_client()
│   │   ├── anthropic/      # Anthropic SDK wrapper
│   │   ├── openai/         # OpenAI SDK wrapper
│   │   └── litellm/        # LiteLLM wrapper (default)
│   ├── infra/              # Storage, broker, secrets, logging ABCs
│   ├── inbound/            # Inbound Update Protocol
│   ├── models/             # Pydantic domain models (17 node types)
│   ├── scoring/            # 7-factor scoring engine
│   ├── skills/             # Skill agent runtime (SKILL.md parser, SkillWorker)
│   ├── state/              # State machine (10 states, guarded transitions)
│   └── triggers/           # Trigger engine (scheduled, inbound, on-demand)
├── tests/                  # pytest test suite (485+ tests)
├── docs/                   # Documentation
│   ├── architecture.md     # Plugin architecture overview
│   ├── llm-providers.md    # How to add an LLM provider
│   ├── channels.md         # How to add a gateway channel
│   ├── db-backends.md      # How to add a database backend
│   ├── api-reference.md    # Gateway API endpoints
│   └── skills-and-agents-roadmap.md
├── docker/                 # Docker Compose local dev stack
├── scripts/                # DB init + seed scripts
├── .claude/                # Claude Code configuration + skills
├── docs/planning/build-plan.md           # 6-phase implementation plan
├── docs/release-notes/CHANGELOG.md
├── docs/governance/documentation-governance.md
└── LICENSE                 # Apache 2.0

Scoring Algorithm

The 7-factor weighted scoring formula prioritizes tasks:

#	Factor	Weight	Description
W1	Timeline Urgency	0.25	Days to deadline, effort slack
W2	Dependency Weight	0.20	Direct + transitive downstream dependents
W3	Critical Path	0.20	On critical path × goal priority multiplier
W4	Blocker Score	0.15	Hard (1.0) / Soft (0.6) blocker elevation
W5	Human Override	0.10	Prioritize (+1.0) / Deprioritize (−0.3) / Snooze (exclude)
W6	Resource Risk	0.05	Reliability, load, risk signals
W7	Constraint Pressure	0.05	Budget/time/resource constraint proximity

Post-multipliers: Critical path P1 goal = 1.5×, P2 = 1.3×, P3 = 1.1×

State Machine

10 task states with guarded transitions:

PENDING → ACTIVE → IN_PROGRESS → COMPLETE (terminal)
                  → BLOCKED → ACTIVE (when blocker resolves)
                  → DELAYED → IN_PROGRESS
                  → NEEDS_REVIEW → COMPLETE | IN_PROGRESS
        → CANCELLED (terminal)
        → SNOOZED → ACTIVE
        → INACTIVE_PENDING → ACTIVE (on predecessor completion)

License

Copyright 2026 Abhishek Gupta. Licensed under the Apache License 2.0.

Build Phases

Phase	Weeks	Status	Focus
Phase 0	1–4	✅ Complete	Core Loop Proof (graph model, scoring, state machine, CLI)
Phase 1	5–12	✅ Complete	Single-User System (gateway, email, triggers, skills, inbound, infra)
Phase 2	13–20	Next	Multi-Channel + Organizations (WhatsApp, Telegram, org workspaces)
Phase 3	21–28	Planned	Multi-User + Security (OAuth 2.0, JWT, IAM, A2A delegation)
Phase 4	29–36	Planned	Visual Interface + Advanced Skills (React UI, calendar, import)
Phase 5	37–48	Planned	Enterprise + Observability (Slack, Teams, GDPR, SOC 2)

See docs/skills-and-agents-roadmap.md for the full skills/agents/channels roadmap.

Contributing

Contributions are welcome! Please read CONTRIBUTING.md and our CODE_OF_CONDUCT.md before opening a pull request.

• License: Apache 2.0
• Security: Report vulnerabilities privately — see SECURITY.md
• Issues: Use GitHub Issues for bugs and feature requests; GitHub Discussions for questions
• PRs: Target the main branch; include tests; ensure pytest tests/ passes; sign off commits with git commit -s (DCO)

Build System

This project is built using the Claude Code multi-agent system:

• Opus — Architecture decisions, planning, complex reasoning, code review
• Sonnet — Code generation, implementation, testing, refactoring
• Haiku — Quick lookups, formatting, simple edits

See .claude/ for agent definitions and custom skills used during the build.

Local Development Tooling

These steps are for contributors inspecting the local stack. Replace all <placeholder> values with the ones you configured in docker/.env.

Run pgAdmin against the local Postgres + AGE container

docker run -d --name pgadmin \
  --network graphclaw_default \
  -e PGADMIN_DEFAULT_EMAIL=<your-email> \
  -e PGADMIN_DEFAULT_PASSWORD=<your-pgadmin-password> \
  -p 5050:80 \
  dpage/pgadmin4

Then open http://localhost:5050/login and connect to the database with:

• Host: db
• Port: 5432
• Username: graphclaw
• Database: graphclaw
• Password: value of DB_PASSWORD from docker/.env

Access the MinIO console for object store inspection

Open http://localhost:9001/login and sign in with:

• Access Key: graphclaw (or value of STORAGE_ACCESS_KEY if overridden)
• Secret Key: value of MINIO_PASSWORD from docker/.env

See docs/how-to/self-host.md for the full self-hosting guide.

License

Apache License 2.0