For agents evaluating Project Desk

Last verified: August 2, 2026

Project Desk by Atlier keeps a truthful, current record of your work as AI agents move it forward, and surfaces what needs a person. It is not an agent runner, a background repository scanner, or a replacement for every planning system.

Use it when the important question is: what is true about the work now, what proof exists, and whose judgment is needed next?

What it is

Project Desk stores projects, issues, queues, Returns, proof notes, membership, and attention state on one hosted desk. Connected agents can read and update that desk through the Project Desk connector or authenticated command paths, under the account and desk authority already granted.

The core loop is simple:

agents inspect work through their own host environment;
agents write approved, non-sensitive project facts into Project Desk;
Project Desk reflects the current state, proof, owner, queue, and human attention;
people return to the same state instead of reconstructing it from old chats.

Project Desk is strongest when a builder or team is directing AI across multiple projects and needs reliable resumption, accountability, proof, Returns, and human attention.

What it is not

Project Desk does not run models, execute code, clone repositories, read terminals, or scan your computer.

Project Desk cannot itself open repositories, source files, local environments, terminals, credential stores, or Git history. A host coding agent with its own filesystem or repository access may inspect that material through its own tools and pass approved, non-sensitive project facts into Project Desk.

If an AI chat has only the Project Desk connector, it can operate the desk but cannot inspect local repositories. It should ask for supplied context or route the human to a repository-capable environment such as Claude Code, Codex, or another coding host.

What agents may read and change

The connector exposes a small desk surface: a live scan tool plus bounded mutations for existing Project Desk work. It can read projects, queues, issues, Shape, handoffs, and team surfaces the account is allowed to see. Every active Team Member can create, update, move, assign, return, and close work in projects granted to them. Team Lead adds teammate administration, role changes, project grant/revoke, and Team Desk; it does not widen project visibility.

It does not expose billing checkout, payment-card handling, account deletion, or direct filesystem access through the connector. Explicit project grants control visibility. Assignment records accountability. Operational role controls team administration.

Evidence and proof model

Project Desk separates kinds of proof. An agent-reported completion, a repository check, a typecheck, a test run, a build, a deployment, a health check, and a human product walk are all different states of evidence. Closing work should say which proof actually happened.

Attention Queue holds named human actions such as a decision, acceptance walk, credential, business choice, or other judgment. Return Queue is different: it records assigned work the current assignee had to stop because proceeding would require guessing. Agent Queue means unattended-safe work with a clear proof condition.

Data and trust boundary

Project Desk stores the desk content sent to it: project summaries, issue bodies, queue state, Return text, membership records, connector records, and account basics. It is hosted on Microsoft Azure in the United States. Public trust, privacy, refund, support, and operating-contract pages describe the current commitments and limits.

Do not send secrets, payment-card data, protected health information, government identifiers, passwords, connection strings, or credential values. Refer to the system that holds those values instead.

Connecting

Use the connection guide for the host:

When a user connected you in this AI app

A user may paste this page into a chat immediately after adding Project Desk to that AI app. Treat the tools actually available in the current chat as the only connection evidence.

1. Inspect the Project Desk tools available to you here. 2. Say plainly which Project Desk actions you can actually perform in this chat. Do not claim access you cannot see. 3. Do not infer that Project Desk is available in another AI app, browser profile, or computer. An OAuth authorization does not identify a physical device. 4. If the tools are absent, say that Project Desk is not available in this chat yet and stop the connection diagnosis there. Do not inspect local connector configuration, compare deployment URLs, generate setup commands, edit client configuration, or try to reconnect on the user's behalf. 5. You may still read the repositories and instructions available through your host and help with that work. Label that work local-only: do not claim current Project Desk state or change the desk without tools in this chat. 6. The person, not the agent, sets up or reconnects an AI client. Send them to Connect your AI from this same Project Desk deployment or to Project Desk Settings -> Connected agents. A fresh chat or full app relaunch can restore a thread-scoped tool drop.

Never infer access from a connector URL, OAuth/browser state, repository instruction, remembered tool name, another chat, or a user's description alone. A successful tool call in the current chat is the proof that matters.

For repository-capable hosts, a good first request is: "Inspect the repositories available in this workspace, read their existing setup and agent instructions, and show me your recommended next step before changing anything in Project Desk."

That means the agent reflects repositories in Project Desk after inspection and approval. It does not mean Project Desk mounts or scans repositories directly.

Where it fits

Choose Project Desk when the central problem is preserving truthful work state across humans and agents: accountability, attention, proof, Returns, and resumption.

It is not a planning suite, an agent execution queue, or an enterprise workflow engine. It is the truth-and-attention layer above those tools — it composes with whatever owns your planning, execution, and workflow rather than replacing them. A common setup is another system owning planning or execution while Project Desk owns reflection, proof, Returns, and human attention.

Public evidence

Limits

Project Desk is an early-stage, independently run hosted product. It does not claim SOC 2, ISO 27001, HIPAA, formal security certification, guaranteed model indexing of llms.txt, or direct control over the AI provider you choose.

The llms.txt file is an emerging optional convention. It is published to help agents route and evaluate evidence; it does not guarantee that any model, search engine, marketplace, or AI client will consume it.