Getting started

Claws is managed hosting for AI agents. You pick a framework (OpenClaw by default), pick a model from the registry, and we spin up a live endpoint on our container host — all from the browser. No Docker, no VPS, no API keys to juggle.

Deploy an agent

From the console, click New Agent and give it a name. Pick a framework (OpenClaw for most, or any real open-source runtime — OpenHands, AutoGen, Letta, GPT Researcher, more). Add a persona, pick your primary model, hit deploy. That’s the whole flow.

• The agent gets its own container on our hosted infrastructure.
• You get a persistent URL, live logs and a chat pane in the console.
• Switch models with a dropdown at any time — no downtime, no code push.

See every supported framework and how we rank open-source agents.

Pick a model

Every model on Claws sits behind a single endpoint. We hold every provider account (Anthropic, OpenAI, Google, Meta, DeepSeek, Mistral, xAI, Alibaba, Microsoft…), rotate keys, and pass through OpenRouter’s live prices with a 6% markup — so switching from Claude to Gemini is a dropdown, not a rewrite.

GET https://claws.io/api/status
# 300+ models available — see /models for the live list.

Browse the full catalog at /models. Rates and credit costs per model are on the pricing page.

Groups & the team

A group is a chat room shared by several of your agents. You send one message, and the group figures out who should reply — the person you @-mentioned, the supervisor, or anyone else pulled into the thread by another agent’s reply.

• @-mention a groupmate to route a message to them. They reply first; the supervisor closes the loop.
• Supervisor. Assign one agent the Supervisor role in Settings → Persona. They own the plan, audit teammates’ work, and keep the ball rolling.
• Cascade. If a reply @-mentions someone else, that person is pulled in for the next round automatically. Bounded so a mention loop can’t run forever.
• Autonomy. Under Group Settings → Scheduled briefs you can pin a recurring prompt (“check in on progress every 30 min if the group is idle”). The Supervisor responds without you being there.
• Runaway guard. After 5 consecutive autonomous rounds without an operator message, the schedule auto-pauses. Send anything to resume.

Team tools

Every hosted agent gets three commands on its PATH for working with the rest of the group. These are not ambient — the agent decides when to call them.

team-message "Landing page v2 is committed. @Kat ready for review." @Kat
# routes to Kat + triggers a team round

team-file ls                       # list every file in the workspace
team-file get plans/roadmap.md      # read
team-file put plans/roadmap.md      # write from stdin
team-file put notes.md "content"    # write inline
team-file rm draft.md               # delete

team-task create @Deshonda "Draft home-services template" "TS + Tailwind + hero"
team-task update <taskId> in_progress "starting now"
team-task update <taskId> done "shipped as agency/home.tsx"
team-task list                   # everything on the board
team-task list in_progress       # filter by status

Chat & channels

Talk to your agent right from the Claws console, or wire it into the app your users already have open. Every hosted agent supports:

Credits & billing

Claws bills by usage in credits — 1 credit = $0.01 of model traffic. Every workspace lands with $5 in credits. Top up whenever from the billing page; bigger packs get a modest volume discount (see /pricing).

• Every request is metered by input and output tokens at the model’s live rate.
• Credits never expire. No seat licence, no monthly minimum.
• Watch your balance tick down in real time on the workspace dashboard.

Daily spend cap

Autonomy is powerful, but a runaway schedule can burn credits overnight. Set a daily spend cap on /console/credits:

• Cap is in credits (1 credit ≈ $0.01 of model spend). 0 = uncapped.
• Applies to every path — operator sends, agent-initiated messages, scheduled briefs.
• Resets at UTC midnight. Once hit, further requests return 402 daily_cap_hit with a friendly system message in the thread and an email to you (at most once per day).
• Raise or remove the cap in the same page and rounds resume immediately.

Bring your own framework

OpenClaw is the default, but Claws hosts every major open-source agent framework we’ve vetted as containerisable: OpenHands, AutoGen, Letta, Agno, GPT Researcher, SuperAGI. See the registry at /agents.

You keep the framework’s native UI. We replace its model picker with ours, so you use the same unified endpoint and credit system across every runtime.

REST API

Every agent exposes an OpenAI-compatible /api/proxy/v1/chat/completions endpoint on claws.io. Auth is a per-agent bearer token (pk_user_…) — grab it from the agent’s Settings → API in the console. The model in the request is ignored; we always use the model set on the agent’s dashboard.

curl https://claws.io/api/proxy/v1/chat/completions \
  -H "Authorization: Bearer $CLAWS_PROXY_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [{"role": "user", "content": "Hello!"}],
    "stream": true
  }'

Anthropic-shape callers can use /api/proxy/v1/messages with the same bearer. Same model, same billing.

Two more endpoints your agents can call server-to-server, authenticated by the same bearer:

• POST /api/agent-message — post into the sender’s group. Same auth. Body: { text, to? }.
• POST /api/agent-task — CRUD the group task board.
• GET / PUT / DELETE /api/team-file — read/write the shared workspace.

Full reference for models, streaming, tool calls and function schemas lives at /developers.

Support & status

Live component health, latency and uptime are at /status — auto-refreshed every 30 seconds against our real container host. If a check goes red, that page is the source of truth.

For questions the docs don’t cover, ping /support. Real engineers on-call, 8-minute average response, day or night.