OpenClaw
Quick Facts
- Category
- Agent runtime (not a model)
- Language
- TypeScript
- Inner loop
- Qwen-Agent · Qwen3-Coder (default) · any OpenAI-compatible endpoint
- Control plane
- Claude Code Slack (operator conversations)
- Security layer
- FrawdBot (behavioral analysis, 21K lines of detection)
- Integrations
- 50+ first-party; plugin system for the rest
- Target hardware
- Mac Mini · MacBook Pro · on-prem Linux
- Deployment time
- 6–12 hours end-to-end
- License
- Open source (see Claw ecosystem on GitHub)
Summary
OpenClaw is the enterprise-tier runtime in the Claw ecosystem — the flagship among six sibling runtimes (NanoClaw, MicroClaw, PicoClaw, ZeroClaw, ExoClaw) that share a common architecture and differ on form factor and host environment. Where OpenClaw fits is the largest envelope: a managed agent that runs on customer-owned Mac hardware, connects to internal tools via 50+ first-party integrations and a plugin system, and is operated through a Slack-based control plane.
The design philosophy is on-premise first. The inference engine and all tool calls live on the customer's hardware. The LLM (Qwen3-Coder by default, but pluggable) is self-hosted. FrawdBot behavioral analysis is bundled in — not an upsell — so insider-threat detection runs against every agent action. Only the operator interface (Claude Code Slack) touches the outside world.
Architecture
- Agent loop — Qwen-Agent wraps the inner tool-use cycle: model call, tool dispatch, result integration, repeat. Uses the MCP protocol for tool extensibility.
- Inference runtime — vLLM or Ollama hosting Qwen3-Coder (or the customer's chosen weights) on Mac Mini M-series hardware. Provider abstraction allows falling back to hosted Anthropic, OpenAI, or Bedrock endpoints.
- Integration layer — 50+ first-party connectors (Slack, Gmail, Google Drive, Linear, GitHub, Notion, etc.). Plugin system for everything else.
- Department loadouts — preset tool bundles, system prompts, and guardrails scoped to a department: SRE, RevOps, Legal, Finance, Support. Faster time-to-value than a blank agent.
- Control plane — Claude Code Slack. Operators approve high-impact actions, view traces, and tune behavior from within Slack threads.
- Observability — OpenTelemetry traces, Langfuse traces, ClickHouse for analytics. Seven metric namespaces exposed to the SRE dashboard. See The Observability Architecture.
- Security — FrawdBot sits inline with every tool call, running behavioral analysis for prompt injection, data exfiltration, and insider threats at machine velocity.
Deployment Model
A typical OpenClaw deployment takes 6–12 hours from first call to production-ready agent:
- Hardware — Mac Mini M-series delivered or provided by the customer. Tailscale tunnel established for management.
- Base install — OpenClaw runtime, Qwen-Agent, chosen LLM weights (Qwen3-Coder default), FrawdBot, observability stack (OTEL + Langfuse + ClickHouse).
- Department loadout — select one or more loadouts. Customize system prompt, tool allowlist, and approval policies.
- Integrations — OAuth or service-account setup for Slack, Gmail, Linear, etc. Plugin installs for anything beyond the 50+ first-party list.
- Control plane — Claude Code Slack channel provisioned. Operators added.
- Ship — agent goes live. SRE dashboard accessible via Tailscale.
When to Choose OpenClaw
OpenClaw is the default pick when any of the following apply: the deployment must stay on customer-owned hardware, the workload benefits from department-specific tool bundles, the customer wants Slack as the agent control plane, or regulatory requirements preclude a cloud-managed agent. For other profiles, another Claw runtime is typically a better fit:
- NanoClaw — when security isolation matters more than integrations (containerized, Anthropic Agent SDK inner loop).
- MicroClaw — when multi-channel chat workflows (Slack + Discord + WhatsApp) with durable session state dominate.
- PicoClaw — IoT, embedded, sensor networks (<10MB RAM, Go).
- ZeroClaw — minimal footprint with maximum LLM provider flexibility (3.4MB binary, Rust).
- ExoClaw — when cloud hosting is acceptable and zero-config matters more than data sovereignty.
Tradeoffs
- On-premise operations cost. Mac Mini hardware, edge inference tuning, Tailscale management, and local observability all live with the customer (and Organized AI as managed-service provider). Cloud-first teams should consider ExoClaw instead.
- Inference ceiling. Edge hardware caps model size. Qwen3-Coder at the scales Mac Minis can host is strong but not frontier. Heavy reasoning workloads route to hosted Claude or DeepSeek-R1 via the provider arbitrage layer.
- Plugin ecosystem maturity. 50+ first-party integrations cover the common cases; the plugin system handles the long tail but may require custom work for niche systems.
- Operator Slack coupling. The control plane assumes Slack. Non-Slack shops can switch to MicroClaw or integrate a different channel, but OpenClaw's defaults are Slack-first.
Related
- Qwen — the LLM family powering the default OpenClaw inner loop.
- Hermes — alternative open-weights fine-tune for system-prompt-driven agent loops.
- Claude — the model family behind Claude Code Slack and the operator-facing control plane.