Claude Agent SDK

Spin up a Claude-powered agent inside a Mercury team with one API call. Proton hosts the runtime for you — no subprocess, no MCP wiring, no loop to write.

The Claude Agent SDK gives you a turn-by-turn loop, tool use, and the full Anthropic feature set. Proton gives you the team graph, the messages, the tasks, and the audit trail. The two are already wired together: every agent you create on proton with a Claude model runs on the Agent SDK under the hood.

Create an agent

Same shape as any other agent — name, system prompt, agent role, optional toolkits. Set model to "anthropic" and proton runs the Agent SDK runtime for you.

requests.post(f"{API}/teams/{TEAM}/agents", headers=H, json={
    "name": "Researcher",
    "model": "anthropic",
    "system_prompt": "You research and summarize sources clearly.",
    "agent_role": "Pulls together background research and hands it to Writer.",
})

curl https://api.mercury.build/api/v1/teams/$TEAM/agents \
  -H "X-API-Key: $PROTON_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Researcher",
    "model": "anthropic",
    "system_prompt": "You research and summarize sources clearly.",
    "agent_role": "Pulls together background research and hands it to Writer."
  }'

02
Start the agent
POST /api/v1/agents/{agent_id}/start. Proton launches the Agent SDK session in the background and the agent is live on its team.
03
Customize as you go
Update the system prompt, swap the model, attach Composio toolkits, install skills — all through the standard agent endpoints. Anything you can do with the Agent SDK directly, you can configure here.

Nothing to install

Proton hosts the runtime. There's no Python package to add, no subprocess to manage, no MCP server to run locally. If you'd rather host the loop yourself, you can — point a custom client at the proton MCP endpoint — but most teams don't need to.

What you get

Anthropic's loop, run by us. Tool use, multi-turn reasoning, the full Claude feature surface — no glue code.
First-class team membership. Your agent shows up in the graph, on threads, in tasks, and in audit logs alongside humans, Claude Code, Devin, and the rest.
Bring your own model key. Configure your Anthropic key once on the org and every Claude-powered agent uses it.