Your AI assistant should be able to see your backlog

You describe a task to Claude or Cursor. It writes the code. Then you describe the same work again: to your tracker, by hand, hours later. The assistant that just did the work has no record it exists.

That gap is why boards rot. Filing is a separate errand, and the errand always loses to the next bug. Close the gap and the board stays true.

Every MCP tracker we could find is a wrapper

The Model Context Protocol is the obvious fix: give the assistant a set of tools and let it read and write the backlog itself. And there are servers for this. Search the MCP directories for project management and you will find Linear, Jira, Trello, Notion, a handful of Task Master variants.

Every one of them is a bridge. A translation layer bolted onto a tool designed for a human clicking a mouse, inheriting that tool's field requirements, its permission model, its idea of what a ticket must contain before it is allowed to exist. The assistant is a second-class client of an API that was never shaped for it.

What we built instead

My Fordyce ships its own server. The tracker and the tool surface are the same product, designed together: thirty-five tools across tasks, projects, releases and roadmap features, named plainly so a model does not have to guess: list_tasks, create_task, move_task, create_release_with_tasks, update_roadmap_feature.

Setup is a URL and a bearer token. No local install, no npm, no Docker:

{
  "mcpServers": {
    "my-fordyce": {
      "url": "https://myfordyce.fordycecg.com/mcp",
      "headers": { "Authorization": "Bearer YOUR_KEY" }
    }
  }
}

That is the whole integration. Generate a key in Settings → API Keys, paste the block into your MCP client config, and the backlog is in the editor.

What you get

• Filing in the flow of work. "File the three follow-ups from this refactor" is one sentence at the end of a session, not a task for Friday. Tickets land within minutes of the thought instead of days.

• An assistant with real context. An agent that can read the board knows what is already in progress, stops proposing work you are halfway through, and notices when a new request contradicts an open ticket.

• Control built in, not bolted on. Lock an assistant to a single project with FORDYCE_PROJECT_ID, let clients prompt before destructive actions, and give every connected editor its own isolated session. Write access with the boundaries drawn where you draw them.

The tool surface is the product

Anyone can bolt a bridge onto an API. Designing the tracker and the tool surface together is what makes an assistant a first-class client: tools named for how a model actually works, a board that answers in one round trip, and nothing lost in translation to a schema built for mouse clicks.

That is why My Fordyce ships its server first-party, and why the board your assistant sees is the same board you see. Point your editor at it and watch the backlog keep itself.