Run Claude Agent SDK in isolated sandboxes. Skills, plugins, connectors, and ready for production. One POST request to launch an agent.
$ curl -X POST $BASE_URL/api/runs \
-H "Authorization: Bearer $API_KEY" \
-d '{"agent_id": "ag_01", "prompt": "Deploy the app"}'Platform
Full Claude Agent SDK agent with tool use, file editing, and bash. Not a wrapper — the real thing running in a sandbox.
Every run spins up a fresh Vercel Sandbox with its own filesystem, network policy, and resource limits.
Inject custom skills into agents. Extend agent capabilities without changing code. Use Claude Cowork plugins.
Connect agents to 900+ external tools via Composio or add your custom MCP servers.
Row-level security, per-tenant API keys, budget controls, and rate limiting. Built for platforms.
Every run stores a transcript, token usage, cost, and duration. Query runs, replay transcripts, track spend.
How it works
Provision isolated workspaces with API keys, budgets, and rate limits.
npm run create-tenant
Set model, tools, permissions, git repo, and MCP connectors.
POST /api/agents
{
"name": "deploy-bot",
"model": "claude-sonnet-4-6",
"composio_toolkits": ["github"]
}POST a prompt, get back a streaming NDJSON response with real-time events.
POST /api/runs
{ "agent_id": "ag_01",
"prompt": "Review the open PRs" }
← {"type":"run_started",...}
← {"type":"assistant",...}
← {"type":"tool_use",...}
← {"type":"result",...}Architecture
/api/agents/:id/runsSSE / NDJSONStop building sandbox orchestration from scratch. AgentPlane gives you a production-ready API for running Claude Agent SDK agents.