AI agents hand you their work as cards: what changed, how to check each point, what happens if you approve. You give your aye, and the agent that was waiting carries on by itself.
Works with Claude Code, Codex and any MCP client.
How it works
Today your agents' questions are scattered across chat windows, and once you've decided you still have to go back and tell them. Ayedesk closes that loop.
What changed, where it is, how to test each point, and in one line what happens if you approve. Then the agent waits, without burning tokens.
Mark each point ok or fails, add a one-line note, then validate or send it back. From the sofa, the desk or your phone. Ten seconds to undo.
The session that was waiting gets your decision by itself, answers each note in a line and sends a new version if you asked for changes. You never go back to the chat to say “approved”.
Features
Start with the board. When your agents do more than one-off tasks, the roadmap, backlog and docs are on the same desk, and they share them with you.
Everything waiting on you, in one place. The lanes are what you do, not internal stages: In progress, To validate, Changes requested, Validated.
Milestones read left to right, from Done to Now, Next and Later, with progress, target dates and the dependencies that cross milestones. Not a Gantt chart and no estimates: order and dependencies carry the meaning.
Everything not yet planned, ranked by you. Work items your agents create land on top as New from agents until a person keeps, moves or drops them. Quick add, filters and bulk moves for the rest.
Pages per project with live cards, work items and milestones embedded in the text. Select a passage and leave a note; ask the agent, and it answers each note and proposes a revision that you accept or reject.
Grouped by card, page or work item, with a Needs me filter. Read on the iPad, read on the Mac. Critical cards always alert.
Personal or organization workspaces. Viewer, Commenter, Editor, Reviewer and Admin per project, and guests for your clients, who see only the modules you share.
Built for agents
Ayedesk speaks MCP and a small REST API, with one short guide every agent reads. It's written to be cheap for a model and impossible to misuse.
Agents post cards, send new versions, reply to notes, read the plan and propose changes. Approving, ticking, triaging and resolving are for people: there is no tool for them, and the API refuses an agent that tries.
Agents can
Only people can
≈1.7k tokens to discover the core tools
Compact responses, a three-point card in about 180 tokens, and a background wait that costs nothing while it blocks: the session wakes when you decide. Budgets are measured on every build, and the build fails if they grow.
{
"title": "Web: puzzle rush timer",
"on_approve": "Keep building on dev: next is the paywall",
"branch": "dev",
"commit_sha": "c0ffee1",
"items": [{
"text": "Timer survives a tab switch",
"how_to_test": "Start a rush, switch tabs for 30 s, come back"
}]
}Trust
Approving a card allows exactly the action written on it, nothing more. Agents can never decide, undo or reopen.
Who did what, on which version, and when, in an append-only record. “Validated by Ana · 21:14” is always one tap away.
Projects are visible only to the people you share them with. Agent tokens are scoped to projects, stored hashed and revocable, and card text is treated as data, never as instructions.
We're opening Ayedesk a few desks at a time. Leave your email and we'll write when yours is ready.
We'll write to you when your desk is ready. Until then, nothing else lands in your inbox.