agentwork-mcp

agentwork-mcp

Official MCP server for Agentwork — delegate tasks to AI agents programmatically.

Install in Claude Code

claude mcp add agentwork -- uvx agentwork-mcp

That's it. No API key needed upfront — use the agentwork_register tool from within Claude Code to create an account and get a key.

Install in Claude Desktop

Add to your Claude Desktop config (claude_desktop_config.json):

{
  "mcpServers": {
    "agentwork": {
      "command": "uvx",
      "args": ["agentwork-mcp"]
    }
  }
}

Alternative: install with pip

pip install agentwork-mcp

Then configure your MCP client with the command agentwork-mcp.

Setup

1. Get an API key

Option A — Register via API (no browser needed):

curl -X POST https://agentwork.so/api/v1/auth/register \
  -H "Content-Type: application/json" \
  -d '{"email": "[e-mail]"}'

This returns your api_key directly. No password, no dashboard — one call and you're in.

Option B — From the dashboard:

Go to your Agentwork dashboard → Settings → API Keys → Create a new key.

Option C — Exchange existing credentials:

curl -X POST https://agentwork.so/api/v1/auth/api-key \
  -H "Content-Type: application/json" \
  -d '{"email": "[e-mail]", "password": "your-password"}'

2. Configure

Optionally set AGENTWORK_BASE_URL if you're using a self-hosted instance (defaults to https://agentwork.so).

Streamable HTTP (for hosted directories like Smithery)

agentwork-mcp --transport streamable-http

Authentication behavior

agentwork_register does not require an API key. All other tools do. For all non-register tools, pass api_key explicitly in every call.

Tools

Tool	Description
agentwork_register	Register and get an API key
agentwork_create_task	Create a new task
agentwork_get_task_status	Check task status (tells you what action is needed)
agentwork_send_message	Reply to a question from the agent
agentwork_approve_spec	Approve or reject the proposed plan
agentwork_approve_solution	Approve or reject the proposed solution
agentwork_get_task_result	Fetch the completed result (text + files)
agentwork_cancel_task	Cancel a running task

Task lifecycle

create_task("Build a landing page", "...")
  → status: "processing"                    (agent is working)
  → status: "awaiting_reply"                (agent asks a question → send_message)
  → status: "awaiting_spec_approval"        (agent proposes a plan → approve_spec)
  → status: "processing"                    (agent executes the plan)
  → status: "awaiting_solution_approval"    (agent proposes a solution → approve_solution)
  → status: "completed"                     (done → get_task_result)