Orchestrating agents over MCP

Skaftor exposes a per-project MCP (Model Context Protocol) server over Streamable HTTP. Coding agents and the skaftor CLI connect to it to pull their work, read the execution context, and report progress — all governed and visible centrally.

Endpoint and authentication

URL:     https://app.skaftor.com/api/mcp/<projectId>
Header:  Authorization: Bearer <your developer token>

Your developer token is personal to you and the project — it sees exactly what you're assigned. Treat it like a password. Regenerating it stops the old one working immediately.

Get your connection details

Open a work order › "Build" › "This machine" and pick your coding environment. Cursor and VS Code get a one-click "Connect & open". For other agents, copy the "MCP" config. The same sheet has a "CLI" row with your skaftor login line.

Claude Code

claude mcp add --transport http skaftorAI "https://app.skaftor.com/api/mcp/<projectId>" \
  --header "Authorization: Bearer <token>"

Run it from your repository, then run /mcp to confirm it connected.

Cursor

Add this to .cursor/mcp.json in your project (or ~/.cursor/mcp.json), then reload Cursor:

{
  "mcpServers": {
    "skaftorAI": {
      "url": "https://app.skaftor.com/api/mcp/<projectId>",
      "headers": { "Authorization": "Bearer <token>" }
    }
  }
}

Codex

Add this to ~/.codex/config.toml, then restart Codex:

[mcp_servers.skaftorAI]
url = "https://app.skaftor.com/api/mcp/<projectId>"
http_headers = { "Authorization" = "Bearer <token>" }

Any other MCP client: point it at the URL with the Authorization header on every request.

Slash commands

The server publishes prompts your agent shows as slash commands, including /next-task, /my-work, /open-work-order, /plan, /start-work-order, /open-pr, /complete-work-order, /build-all and /ask-about-work-order. Start with /open-work-order <code>.

What the agent can do

• Find its work — next_work_order, my_work_orders, list_work_orders, get_work_order.
• Plan before coding — plan_work_order returns the repository, files, reuse and impact; get_execution_manifest returns everything in one payload.
• Find the checkout — resolve_workspace locates the work order's repository on your machine.
• Report progress — start_work_order, sync_execution (commits), ask_about_work_order.
• Deliver — prepare_pull_request and complete_work_order attach the pull request and move it to review.
• Workstations — launch_workstation, list_workstations, transition_workstation.