The control plane for governing, securing, observing, auditing, and operating AI agents and MCP servers.
Kubernetes · ServiceNow · Splunk - for AI Agents.
ClawForge is not another agent framework. It is the control plane for managing, governing, securing, observing, auditing, and operating AI agents, MCP servers, workflows, tools, models, and enterprise integrations - built for government entities, municipalities, and enterprise IT, security, and AI platform teams.
| Product | Role |
|---|---|
| Hermes | Learns |
| OpenClaw | Executes |
| Paperclip | AI company OS |
| ClawForge | Enterprise / Government Agent Control Plane |
Beneath the control plane sits a high-performance, local-first agent runtime
written entirely in Rust - a Rust implementation of the
OpenClaw topology that orchestrates autonomous agents over
a central WebSocket gateway, across channels (WhatsApp, Telegram, Discord, Slack),
with tools, sandboxing, and memory. The clawforge-controlplane crate wraps that
runtime with the registry, governance, security, observability, and compliance
layers an organisation needs to run agents safely at scale.
ClawForge is an enterprise-grade AI agent control plane for governing, securing, monitoring, auditing, and operating AI agents and MCP servers across government and enterprise environments.
ClawForge governs your AI agents; it does not replace them. Build agents with whatever framework, model, or MCP server you like. ClawForge sits above them as the control plane that decides which agents may exist, what they are allowed to do, who approved them, what they actually did, and whether that was compliant.
- It is: a control plane for AI agents (registry, approval governance, per-action security gateway, fleet observability and audit, MCP and integration governance, compliance).
- It is not: an agent framework or SDK, an LLM, a turnkey "AI dev team", or a replacement for your runtime.
Operators and governance teams drive the control plane; the control plane authorises and governs the agent runtime; the runtime streams execution events back for observability and audit. See docs/architecture.md and docs/diagrams.md.
The clawforge-controlplane crate adds the layers an organisation needs to run
agents safely at scale. Each is a self-contained, SQLite-backed, fully tested
domain module:
| Capability | What it does | Docs |
|---|---|---|
| Agent Registry | Single source of truth for every agent (owner, tools, MCP, data access, risk, lifecycle) | registry.md |
| Governance Engine | Human approval workflow with department ownership, change history, and audit | governance.md |
| Observability | Execution events → task / cost / latency / failure / risk metrics, per-agent and fleet-wide | observability.md |
| Security Gateway | Pre-execution checks on every action (tool / MCP / model / data / budget / approval) + risk score | security-gateway.md |
| MCP Governance | Registry, approval, health, and usage tracking for MCP servers | mcp-governance.md |
| Agent Marketplace | Verified, reusable internal agent templates with compliance badges | marketplace.md |
| Enterprise Integrations | Governed connectors (DBs, SSO, GIS, ITSM) - credentials referenced, never stored | enterprise-integrations.md |
| Government Compliance | PII classification, retention, approval chains, audit evidence, reporting (UAE PDPL-aware) | government-compliance.md |
Requires Rust ≥ 1.80 (and Node ≥ 20 only for the web dashboard).
git clone https://github.com/YASSERRMD/clawforge.git
cd clawforge
# Build & test the control plane
cargo build -p clawforge-controlplane
cargo test -p clawforge-controlplane # 82 tests
# Run the end-to-end control-plane demo
cargo run -p clawforge-controlplane --example demoThe demo walks a single agent through the whole control plane - marketplace install → MCP approval → governance → security gateway → observability → compliance report - in memory. See docs/demo.md.
# Set a key for any provider(s) you want; the runtime registers each one present.
export ANTHROPIC_API_KEY="sk-ant-..." # or OPENAI_API_KEY, DEEPSEEK_API_KEY, OPENROUTER_API_KEY, ...
cargo run -p clawforge-cli -- serve --port 3000 # local-first gateway
cd frontend && npm install && npm run dev # dashboard (separate terminal)
# or: docker-compose up --buildClawForge is multi-provider: OpenAI, Anthropic, Google Gemini, Mistral, xAI, Groq, OpenRouter, local Ollama, and the major Chinese providers (DeepSeek, Qwen, Zhipu GLM, Moonshot/Kimi, Baidu ERNIE, MiniMax, Tencent Hunyuan, 01.AI Yi, StepFun, Baichuan, iFlytek, SenseTime). See docs/model-providers.md.
Configuration is environment-driven; see .env.example and docs/installation.md.
- Start here: Product positioning · Architecture · Diagrams · Installation · Model providers · Demo
- Use cases: Overview · Government municipality · Enterprise IT
- Operations: UAE PDPL note · Security disclaimer · Roadmap · Limitations
- Contributing: Developer guide · CONTRIBUTING.md
Full index: docs/README.md.
ClawForge's control plane governs a complete, local-first agent runtime (Rust workspace):
clawforge-core- central schemas (AgentSpec,Message,Event).clawforge-gateway/clawforge-daemon- Tokio WebSocket control plane for sessions, tools, and events.clawforge-planner/clawforge-executor- LLM provider integrations (OpenRouter, Ollama) and sandboxed action execution.clawforge-channels- deep adapters for Telegram, Discord, Slack, LINE, iMessage, and WhatsApp.clawforge-plugins/clawforge-browser/clawforge-understanding- WASM plugins, CDP browser automation, and OCR/STT/PDF media pipelines.clawforge-memory/clawforge-supervisor- vector memory for RAG and SQLite run-state persistence.
ClawForge connects to real messaging surfaces - treat inbound messages, tool output, and MCP responses as untrusted input. The Security Gateway gates capabilities (which tool / MCP / model / data an action may use); keep untrusted execution sandboxed (the runtime supports Docker isolation). Secrets are never stored - integrations hold credential references only. Read the full security disclaimer and limitations before relying on ClawForge.
Contributions are welcome. Please read the developer guide
and CONTRIBUTING.md - work in atomic, well-described commits,
run cargo test -p clawforge-controlplane before opening a pull request, and keep
documentation in step with code.
MIT.