Docs

Agents

Create OpenClaw, Hermes, NemoClaw, or OpenBot agents that call the live gateway and spend this workspace's quota.

Dashboard → Agents or the gateway Agents API lets a workspace stand up its own agent runtime. Agents is a $20/month add-on (included on Enterprise and Family Max). Pick OpenClaw, Hermes, NemoClaw, or OpenBot, choose a catalog model, and spin it up. Completions use a dedicated API key, so spend hits the same prepaid quota as Playground and /v1/chat/completions.

Pricing

The add-on unlocks the hosted runtimes. It is a separate Stripe subscription from Pro/Team, so canceling Agents does not cancel the seat plan. Token usage still bills workspace quota. Subscribe from Agents or Billing.

Runtimes

RuntimeBest for
OpenClawMulti-channel gateway agent (Slack, Discord, Telegram)
HermesOne self-improving agent with durable memory
NemoClawNVIDIA's sandboxed OpenClaw for a tighter production runtime
OpenBotCopilotKit coworker with a browser, /workspace files, and take-the-wheel

What create / start does

  1. Writes a workspace_agents row for the chosen runtime and model.
  2. Mints a scoped API key (Agent · <name>) allowed to call that model.
  3. Probes GET /v1/models with that key. If the gateway is down, the agent is failed with the real error.
  4. Seeds runtime skills (OpenClaw channels, Hermes memory/skills, NemoClaw sandbox).
  5. Marks the agent running. Chat runs a tool loop against the live gateway, so spend hits Usage and Billing.

Hosted runtime

Dashboard POST /api/agents/:id/chat and gateway POST /v1/agents/:id/chat run the same loop (tools + memory + outbox + OpenBot computer). The dashboard streams events; the API returns JSON { reply, events, workspace, agent }. Stop pauses spend. Delete removes the agent.

OpenBot opens from Dashboard → Agents → OpenBot inside the usual dashboard chrome (sidebar, search, mobile nav). The coworker shell adds a retractable channel rail, “Start a new channel”, Ask anything, and explore cards. A channel is chat plus the live screen. You can take the wheel, type a secret that never enters the transcript, and hand the browser back. Screenshots and control go through /api/agents/:id/computer.

OpenBot is the MIT coworker stack from CopilotKit/openbot, not CopilotKit Intelligence. The computer is apps/openbot-computer (Playwright Chromium, /workspace, take-the-wheel). The supervisor (apps/openbot-supervisor) starts one container per agent so two OpenBots never share a browser or /workspace. Point the gateway at it with OPENBOT_SUPERVISOR_URL and OPENBOT_SUPERVISOR_TOKEN. docker compose up -d openbot-supervisor builds the image and starts the supervisor.

A shared single computer (OPENBOT_COMPUTER_URL) is the laptop fallback. Without either, navigate/read/files still run in-process so CLI and SDK keep working. Set OPENBOT_COMPUTER_RUNTIME=runsc when the host has gVisor.

Computer actions are fail-closed: public http(s) only, /workspace paths only, snapshot-then-click, and computer_request_help on a login wall. PATCH /v1/agents/:id with { "computerControl": "take" } lets a person take the wheel. POST /v1/agents/:id/ag-ui speaks AG-UI.

CLI and SDK

Same objects from a workspace API key — see SDK & CLI, CLI, and TypeScript SDK.

bash
bun run od -- agents create --name Desk --runtime openbot --model llama-3.1-8b-instruct bun run od -- agents chat --id "$AGENT_ID" --message "Open https://example.com and summarize the page."