Docs

Create hosted OpenClaw, Hermes, NemoClaw, or OpenBot runtimes and chat with them over the same API key as the rest of the gateway.

Agents

Workspace agents are first-class API objects. Create one, it boots against the live gateway, then POST /v1/agents/:id/chat runs the runtime tool loop. Completions use a dedicated agent key, so spend hits the same prepaid quota as Playground and /v1/chat/completions.

Agents is a $20/month add-on (included on Enterprise and Family Max). Without it these routes return 402 with code: addon_required.

GET    /v1/agents
POST   /v1/agents
GET    /v1/agents/:id
PATCH  /v1/agents/:id
POST   /v1/agents/:id/chat
POST   /v1/agents/:id/ag-ui
POST   /v1/agents/:id/restore
DELETE /v1/agents/:id

These are the same operations as the dashboard Agents / OpenBot desk: list, spawn a coworker or Leaderbot, start/stop, chat, and 7-day soft delete. Computer CDP (screenshot, click, secret field) stays on the dashboard cookie routes and the existing computer token — it is not exposed on /v1/agents/:id/computer.

Create and boot

bash
export OPENDOOR_API_KEY=opd_…
export OPENDOOR_BASE_URL=https://opendoor-gcp.web.app

curl "$OPENDOOR_BASE_URL/v1/agents" \
  -H "Authorization: Bearer $OPENDOOR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Desk",
    "runtime": "openbot",
    "modelId": "gemma-4-26b-a4b-it"
  }'

runtime must be openclaw, hermes, nemoclaw, or openbot. modelId is a live catalog id. Optional systemPrompt overrides the runtime default. Optional kind is leader or coworker. kind: "leader" (or the name Leaderbot) is idempotent: if Leaderbot already exists the existing row is returned; if it is in the 7-day recovery window it is restored and booted.

Create mints a scoped API key, probes GET /v1/models, seeds runtime skills, and marks the agent running. If the gateway is unreachable the row is still created with status: "failed" — call PATCH with { "status": "running" } to retry.

Plan seat caps apply: maxApiKeys is the bot cap and maxActiveDeployments is the concurrent running cap. Hitting either returns 402 with code: at_bot_cap or at_concurrent_cap.

GET /v1/agents also returns deleted (recoverable rows), addon, and capacity. GET /v1/agents/:id includes recent messages.

The response is a public agent object: id, runtime, kind, model, status, and a workspace snapshot (memory, skills, outbox, computer). Secrets are never returned.

Chat

bash
curl "$OPENDOOR_BASE_URL/v1/agents/$AGENT_ID/chat" \
  -H "Authorization: Bearer $OPENDOOR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"message":"Open https://example.com and summarize the page."}'

Returns { object: "agent.turn", reply, events, workspace, agent }. The agent must be running (409 otherwise). Tool events are the same loop the dashboard Agents page uses.

OpenBot computer tools

OpenBot (runtime: "openbot") uses the MIT OpenBot computer contract. Set OPENBOT_SUPERVISOR_URL + OPENBOT_SUPERVISOR_TOKEN so each agent gets its own Chromium container (apps/openbot-supervisor launching apps/openbot-computer). OPENBOT_COMPUTER_URL is a shared-computer fallback.

The dashboard coworker desk at /dashboard/agents/:id is the OpenBot UI: live screen, take-the-wheel, and a secret field that never enters the transcript. It calls /api/agents/:id/computer/screenshot and /control.

Tool	What it does
`computer_navigate`	Open a public `http(s)` page
`computer_read`	Read the current page as text
`computer_snapshot`	List fields/buttons with refs (`e1`…)
`computer_click` / `computer_type` / `computer_key` / `computer_scroll`	Act using `ref` + `snapshotId`
`computer_screenshot`	PNG for the person watching
`computer_list_files` / `computer_read_file` / `computer_write_file`	`/workspace` files
`computer_request_help`	Stop at a login, 2FA, or captcha wall
`computer_request_secret`	One masked value into a field; the Bot never sees it
`render_component`	Structured UI the dashboard can show

computer_read_page and request_help remain aliases. Every computer action is decided and audited before it runs. POST /v1/agents/:id/ag-ui streams the same turn as AG-UI.

Start, stop, take the wheel

bash
curl -X PATCH "$OPENDOOR_BASE_URL/v1/agents/$AGENT_ID" \
  -H "Authorization: Bearer $OPENDOOR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"status":"stopped"}'

bash
curl -X PATCH "$OPENDOOR_BASE_URL/v1/agents/$AGENT_ID" \
  -H "Authorization: Bearer $OPENDOOR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"computerControl":"take"}'

status is running or stopped. computerControl is take or release (OpenBot). You can send both in one PATCH. Optional name, systemPrompt, and modelId update the row the same way as the dashboard.

Restore (7-day soft delete)

bash
curl -X POST "$OPENDOOR_BASE_URL/v1/agents/$AGENT_ID/restore" \
  -H "Authorization: Bearer $OPENDOOR_API_KEY"

DELETE stops the agent and keeps the row for 7 days (recoverUntil is in the response). Restore returns the agent as stopped — call PATCH with { "status": "running" } to attach the computer again. After the window, purge is permanent.

CLI and SDK

bash
export OPENDOOR_BASE_URL=https://opendoor-gcp.web.app
bun run od -- agents create --name Desk --runtime openbot --model gemma-4-26b-a4b-it --kind coworker
bun run od -- agents chat --id "$AGENT_ID" --message "Hello"
bun run od -- agents computer --id "$AGENT_ID" --take
bun run od -- agents restore --id "$AGENT_ID"

ts
const agent = await client.agents.create({
  name: "Desk",
  runtime: "openbot",
  modelId: "gemma-4-26b-a4b-it",
  kind: "coworker",
});
const turn = await client.agents.chat(agent.id, { message: "Hello" });
await client.agents.computer(agent.id, "take");
await client.agents.restore(agent.id);

Marketing SDK page — how to get a key and work from the terminal
CLI
TypeScript SDK
Agents in the product
Platform API

CLI