Skip to guide

Project Desk Guide

Choose a working setup, then follow the next useful path.

Guide v1.0.3 · updated 2026-08-27

Choose your working setup

Project Desk works with the workspace your AI client already has. Choose the path that matches how you want to work today.

One repository at a time

Start your agent inside one repository. It can read and change files only in that checkout.

Working in one repository

Across a portfolio

Start your agent in a parent folder that contains the repositories you want it to use. It can deliberately move between those checkouts and complete scoped Agent Queue work across projects, one repository at a time.

Working across a portfolio

Your AI client controls which local folders the agent can access. Project Desk never searches the rest of your computer.

Start here — your first 10 minutes

Project Desk is your work desk for building software with AI. Your chat and connected agents move the work; the desk keeps a truthful picture of it. Here is the shortest path to a useful first action.

  1. Connect an AI client once to your desk and confirm it can see your work.
  2. Open a project you can see and take or create one piece of work.
One repository at a time

Start your agent inside a single checkout. Your whole desk is still visible — the connection belongs to your Project Desk account, not to a repository — but the agent can only read and change files in the repository it was started in.

That is usually what you want while you are working one project, and it has one consequence worth knowing before you rely on it: a chat that can reach one repository can only finish work in that repository. If you queue scoped Agent Queue tickets across several projects and expect an agent to work through them unattended, this setup will finish only the ones belonging to the checkout it is in. The rest will sit.

To run a batch that spans projects, start a chat in the folder that contains those repositories instead.

Across a portfolio

Start your agent in a parent folder that holds the repositories you want it to use. One chat can then move between those checkouts deliberately and finish scoped Agent Queue work across several projects in a single run. This is the setup for handing an agent a batch of tickets and letting it work through them unattended.

The trade is reach: the agent can open any repository inside that folder. Your AI client controls which folders it may access, and Project Desk never searches the rest of your computer — but inside that folder, scope is yours to keep.

Name the project you mean when you hand off work, and read the receipt to confirm the change landed where you intended.

The mental model

Four ideas cover almost everything. A project holds issues (the units of work). Your desk is the current picture of that work, and queues show what is moving, waiting, or needs you. At any moment a piece of work has an owner — who has the ball.

The facts that matter when you work solo with connected agents:

  • Project Desk reflects your work; it does not orchestrate your agents. The desk mirrors the state your chat and connected agents move — it is not a controller that runs them.
  • Project Desk is not a scoreboard. It shows where work sits and who is accountable — never per-person productivity metrics, rankings, throughput, or return-rate aggregates.
  • Queues read flow before horizon: Active and Attention work is never pushed behind Parked work.
  • Every action returns a receipt that names the destination, the recipient, and the resulting item, so you never have to infer success from a version number.
Connect an AI client

Project Desk works through the AI client you already use, such as Claude Desktop, Claude Code, or an OpenAI Apps client. Connect it once: the connection belongs to your Project Desk account, not to one repository. Open your client's connector or integration settings, add Project Desk as a connector, and sign in when prompted. You should then see Project Desk among the client's available tools. One connection reaches every desk your account can use, so if you have more than one, your agent names the desk it means for each piece of work rather than Project Desk remembering a current one for you. When you begin work in a repository, scan your desk and choose or create the matching project explicitly. To confirm the connection is working, ask your agent to scan your desk once: a short read of your projects and queues comes back, which means the connection is live. Your agent handles the technical protocol behind this; see For agents.

Work recipes

Short, repeatable ways to do the common things well.

Creating a useful issue

Name the outcome you want, the context your connected agent needs to start, any constraints that bind the work, and what proof will show it is done. A vague title like "fix the thing" is the common miss: it forces the next step to guess.

Put unattended-safe work in the Agent Queue

The Agent Queue is for scoped, well-specified work an agent can finish without a human watching. Give it a clear outcome, firm boundaries, and a definition of done. Dropping ambiguous or judgment-heavy work there is the mistake: it needs a person, not a queue.

Ask a human on the Attention Desk

When you need a person's judgment or a decision, raise it on the Attention Desk with enough context to act on right away: what you need, why, and what you have already tried. A bare "please look" with no context is the mistake.

Close work with shipped evidence

Close work by naming what actually shipped and any human gate that still stands between it and done. Closing with no evidence, or leaving a remaining human step unnamed, is the mistake: the next person cannot tell finished from abandoned.

Read a refusal receipt as your next step

A refusal receipt names the specific next step, for example grant access first, or reassign before revoking, not just that something failed. Read it and take that step. Retrying the same action blindly is the mistake.

Connected agents and product boundaries

What Project Desk is — and, just as importantly, what it is not — when you point an agent at it.

  • Project Desk stores the desk text sent through its surfaces. It does not open your repositories, terminals, credentials, or local files.

You may notice a versioned PROJECT_DESK.md appear in a connected repository. Your agent writes and updates it; you do not hand-edit it. If AGENTS.md or CLAUDE.md is shared, symlinked, or governed elsewhere, your agent leaves it alone and keeps the Project Desk block in PROJECT_DESK.md.

For the agent's own technical protocol, see /for-agents.

When something looks wrong

Start from what you observed. Diagnose before you change anything.

What does Project Desk store, and what can it not reach?

Project Desk keeps only the desk text you send through its surfaces. It has no way to open your repositories, terminals, credentials, or local files, so using it exposes nothing on your machine. It reflects your work; it does not reach into where that work lives.

My agent connected, but nothing shows a verified use yet.

Connecting is not the same as making a first call. Ask your agent to scan your desk once, and the use will register. For the technical protocol, see /for-agents.

The Project Desk tools disappeared from my AI client.

Reconnect the same connector rather than creating a renamed duplicate, and check that you are signed in as the right account. See /for-agents for the connection protocol.

I seem to be signed in as the wrong identity, or my session is stale.

Sign out and back in, then confirm which email you are signed in with. One person maps to one account by verified email, so the correct email restores the correct desk.

Work is showing up in the wrong queue.

Move it to the right queue. Queues read flow before horizon, so Active and Attention work is never pushed behind Parked; the piece you moved settles into flow order automatically.

What is the difference between an Attention and an assigned-work Return?

An Attention asks a human to look at something. An assigned-work Return hands work you were assigned, but cannot safely finish, back into the shared Return Queue with a clear reason.

How do the Agent Queue and Attention Queue relate to issue status?

They work oppositely. The Agent Queue is a status: an issue is agent-ready exactly while its status is todo, so moving status to or from todo moves it in or out of the Agent Queue (a walk or proof, or a todo issue carrying an open Attention, appears in the Attention Queue instead). The Attention Queue is not a status: an issue is there because it has an open Attention, whatever its status, so changing it to Backlog or Parked never takes it off. Resolve the Attention to take it off, and the issue keeps its status. Closing the issue resolves its Attentions automatically.

How do I avoid creating duplicate Attention or duplicate Return records?

Check the queue for an existing ask before you create a new one. If one is already there, reuse or update it rather than opening a second.

A revoke was refused because work would be stranded.

The refusal is protecting the work, not blocking you. Reassign or open-handoff the live work first, and the revoke will then succeed.

A receipt said success, but the surface I expected did not change.

Read the receipt: it names the real destination, recipient, and resulting item. The change most likely landed where the receipt says, so do not retry blindly. If the destination still looks wrong, contact /support.

How do I export my data or contact support?

Export from Settings, under Your data. To reach a human, use /support; we aim to reply within a few business days.

What an empty surface means

An empty desk surface is usually healthy, not broken. Here is what each one means and the next legitimate step.

Attention Desk — empty

Nothing needs your attention right now, which is a good sign rather than a problem. When you are ready for more, pick up Active work or open a project from your Project Desk.

Agent Queue — empty

No work is queued for an agent yet. The Agent Queue holds tasks an agent can safely finish on its own, without someone watching over it. When you have one ready, add a scoped, clearly specified task here.

Return Queue — empty

No assigned work has been returned. A connected agent sends a Return here when it cannot safely finish something it was given, so an empty queue means nothing is waiting on you.

Archive Desk — empty

Nothing has shipped yet. When you close work as done it comes to rest here, so this is where you will find what you have completed.

Still stuck? Contact support. We aim to reply within a few business days.

Project Desk by Atlier · Desk · For agents · Support · Trust · Privacy