Connect your agent

Your factory speaks the Model Context Protocol. Any MCP-aware client (Claude Code, Claude Desktop, Cursor, a custom AI SDK app) can read and mutate your factory's tasks, documents, runs, and logs through one server, with no per-client integration.

The server lives at https://<your-platform-domain>/mcp and uses OAuth 2.1 with dynamic client registration: the first time your client calls a tool, a browser opens to the consent page where you sign in and pick an organization.

Claude Code

claude mcp add --transport http sfab https://<your-platform-domain>/mcp

Claude Desktop

Add to your MCP config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

{
  "mcpServers": {
    "sfab": { "url": "https://<your-platform-domain>/mcp" }
  }
}

Clients that only speak stdio

Use the mcp-remote proxy. It runs locally and handles the OAuth flow in your browser:

npx -y mcp-remote https://<your-platform-domain>/mcp

One connection = one organization

At consent time the connection is bound to the organization you pick, and every tool call resolves to it. Tools never take an organization parameter, so a leaked token can never operate outside its org. To work with a second org, register the server a second time under a different name and pick the other org when prompted.

Access tokens last 24 hours and refresh automatically for two years; after that the client re-runs the consent flow.

What your agent can do

The MCP server exposes the full factory tool surface (tasks, comments, documents, repos, runs, activity, logs, sandboxes, waits, org status, and more). Every tool advertises its input schema over MCP. Families include, among others:

Clients discover exact tool names and required fields from the live schema — treat the table as a map of surfaces, not an exhaustive catalog.