OpenClaw

A single Gateway process is the local control plane. It owns sessions, tools, events and channel connections, and the Control UI, CLI, TUI and companion apps all talk to it over WebSocket. By default it runs one agent (main); adding agents with openclaw agents add gives each one a separate workspace, auth profiles and SQLite session history, and bindings decide which chat account reaches which agent.

Delegation inside a run is supervisor-style. The main agent calls sessions_spawn to start a sub-agent in its own session; the call returns immediately and the child announces its result back to the requester. Children can spawn further children up to a configurable depth (5 by default), with per-session caps on active children. For large fan-out, Swarm lets a Code Mode script launch many collector children and await their structured results with ordinary JavaScript control flow.

The agent loop itself is pluggable: turns can run on OpenClaw's built-in runtime, on the Codex app-server through the official codex plugin, or on external harnesses such as Claude Code, Gemini CLI or OpenCode via ACP. Long-term memory lives in Markdown files in the workspace, with a SQLite-backed builtin search engine; sub-agents receive only AGENTS.md, not the parent's persona or memory files.

Host command execution passes through exec approvals: policy, an allowlist and an optional prompt must all agree before a command runs, and approval prompts can be answered in the Control UI or in chat apps such as Slack, Telegram and Discord. MCP servers can be set to approve, prompt or auto per server. The main agent can call ask_user to pause and ask one to three structured questions (sub-agents do not get this tool). A human can redirect a running turn with /steer, stop a parent run together with its children, and inspect child runs with /subagents. Unknown senders on DM-capable channels must be approved with openclaw pairing approve before they can talk to the assistant.

npm install -g openclaw@latest --allow-scripts=openclaw

# needs Node 24.16+ or 26.1+; npm 11.16+ for --allow-scripts
npm install -g openclaw@latest --allow-scripts=openclaw

# onboarding checks model access, creates the workspace, installs the Gateway service
openclaw onboard --install-daemon
openclaw gateway status

# a second isolated agent with its own workspace
openclaw agents add research --workspace ~/openclaw-research

# give agents an MCP server, then check it answers
openclaw mcp add memory --command npx --arg -y --arg @modelcontextprotocol/server-memory
openclaw mcp doctor memory --probe

# open the Control UI and chat
openclaw dashboard

• Requires Node 24.16+ or 26.1+ (Node 26 recommended). The hosted installer (curl -fsSL https://openclaw.ai/install.sh | bash) provisions Node if it is missing.
• --allow-scripts=openclaw is for npm 12 or npm 11.16+; on npm 11.15 and earlier run the same command without it. npm approve-scripts openclaw does not work for a global install.
• Plain npm install in a source checkout is not supported; the repository is a pnpm workspace.
• Tools run on the host for the main session unless sandboxing is configured; read the security and sandboxing guides before exposing the Gateway or adding other users.
• Several plugin SDK compatibility aliases have removal targets of 2026-09-30 and 2026-10-01 (v2026.9.7 release notes).

Yes. Agents can use tools from configured MCP servers, openclaw mcp serve exposes channel conversations to MCP clients, and a bundled A2A channel plugin accepts A2A 1.0 JSON-RPC tasks and can message peer agents. AG-UI support is not documented.

Yes. The README says the project has no paid tier, hosted service or token; you pay only the model providers you configure. The LICENSE file contains the MIT license text.

Yes, in two ways: several isolated agents in one Gateway routed by channel bindings, and sub-agents spawned by a running agent with sessions_spawn or fanned out with Swarm.

Both are self-hosted assistants reachable from chat apps. OpenClaw is TypeScript and run by a foundation with no paid tier; Hermes Agent is Python and maintained by Nous Research. OpenClaw's docs include a source-by-source comparison page written from OpenClaw's side.

Yes. openclaw attach launches Claude Code with a scoped Gateway MCP grant, the official codex plugin runs turns on the Codex app-server, and ACP sessions can run Claude Code, Codex, Gemini CLI or OpenCode.