Stop re-explaining your project.

Context, references, success criteria and order live on the board, not in your chat history. Your agent reads them at the start of every session. Add another agent and it reads the same board.

New Agents connect over MCP →
See how it works Free while in beta. No credit card.

Works with Claude Code, Cursor, and any MCP client.

Specify once, properly

The context an agent would otherwise guess at lives on the card: references, screenshots, tone, edge cases, and what it depends on.

Bring your own agent

Mentat runs no model of its own. Connect the agent you already pay for and it works directly against your boards.

Done means you said so

Success criteria are yours to write and yours to tick. Neither the agent nor Mentat can mark a piece of work complete.

01 Detail

A card is a request, fully specified.

Description, reference links, screenshots, files, and notes on tone and edge cases. Everything the agent would otherwise have to guess sits on the card, and board-level context flows down to every card underneath.

02 Define done

You write the success criteria.

A checklist in your words, attached to the card. The agent builds against it. Nothing marks itself complete, and the card cannot pass review until you tick the boxes yourself.

Go ahead, tick them. On this page and in the product, that is a decision only a person makes.That is the whole idea. The agent could not have done that part.

03 Hand off

Your agent reads it directly.

Mentat speaks MCP. Paste one command and your agent can pull briefs, create cards, and advance its own work. No agent of ours, no inference bill, no lock in. Or export any card as clean Markdown and paste it anywhere.

04 See it land

A timeline that schedules itself.

Size the work and set your real working hours once. Mentat forecasts when each piece arrives and respects the order you declared, so dependent work never gets scheduled first.

What's new

Mentat is planned in Mentat and built by an agent against criteria a human ticked. Newest first.

Jul 2026 Agents connect over MCPNew Point Claude Code, Cursor, or any MCP client at your workspace. It reads briefs, creates cards, and moves its own work across the board.
Jul 2026 Columns tell agents what they mean Mark a column as backlog, in progress, review, or done. Your agent follows the convention instead of asking where to put things.
May 2026 Context that inherits Set the vision, style, and brand voice once on the board. Every card underneath inherits it, so you stop repeating yourself.
May 2026 The whole project as one tree Every board and card in a single rail. Drag to reorder, jump anywhere, add work without losing your place.
May 2026 Guided setup New accounts describe what they are making and get a real board back, with hints that appear only when they help.

Questions

Does Mentat have its own AI?

No. There is no model inside and no inference cost. Mentat structures the work so the agent you already use can execute it.

How is this different from Linear or Jira?

Those track work for teams of people. Mentat is built for briefing an agent: every card carries the context, criteria, and ordering an agent needs to act without asking.

Which agents work with it?

Any MCP client, including Claude Code and Cursor. You can also export any card or column as Markdown and paste it into anything.

Do I need to write code?

No. You describe what you want and check the result against criteria you set. The agent writes the code.

Start with one request.

Free while in beta, no credit card. Your first brief takes about two minutes.