Agents

Agents do the work. People decide.

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: there is simply no tool to approve.

tokens to list the core tools: 1,709
tokens for the whole agent guide: 598
tokens to post a three-check card: 180
tokens while waiting for you: 0

In your agent's session

It asks, waits for your aye, then carries on.

Before it does something that needs a person, your agent posts a card to Ayedesk: what changed, what to check and what approval allows. Then it stops and waits, without spending tokens.

You tick the checks and decide, from the Mac, the iPad or your phone. The session that was waiting picks up your decision by itself and does exactly what the card said. If you asked for changes, it fixes them and sends the same card again.

See the whole loop

Agent session · pricing-site

YouRewrite the pricing page copy, then publish it on the dev site.

The new copy is ready. Before I publish it, I've asked you to check it on Ayedesk.

#105Pricing page copy

3 checks for you

If approved: publish the copy on the dev site

Waiting for your aye0 tokens while you decide

Aye from Jordi3 of 3 checks OK

Approved. Publishing the copy on the dev site now.

Integrations

Bring the agents you already use.

Each integration is a thin wrapper around the same guide and the same tools; none of them adds a permission. They connect with an agent token you create in Settings, scoped to the projects you choose, shown once and revocable. Agent kit v1.2.0.0.

Works with

Claude Code
Codex CLI
Any MCP client
claude.ai and ChatGPTsoon

Claude Code

Available

One plugin with everything: the agent guide, the connection to your desk and the helper that waits.

Waits in the background and wakes the session when you decide, with no tokens spent in between.
Loads Ayedesk's tools only when the agent needs them.

Codex CLI

Available

Connect Codex to your desk and add a short note to your project, so it knows when to ask for your aye.

Starts with the tools for cards and decisions; the roadmap, docs and files are one step away.
After posting, Codex ends its turn and picks up your decision from its inbox.

Any MCP client

Available

Any agent that speaks MCP can connect with an agent token, with the agent guide as its instructions.

Choose how much it sees: only cards and decisions, or the roadmap, docs and files too.
Agents that can run commands wait in the background; the others pick decisions up from their inbox.

claude.ai and ChatGPT

Coming soon

Hosted connectors need a public endpoint with OAuth sign-in, which we're building now. A connected app stays an agent: it signs in with your consent and never inherits your right to decide.

Sign-in with your consent, scoped to your projects.
Decisions come back through the inbox when you return.

Set up in three steps

Create an agent token

In Settings, choose the projects it can reach. It's shown once, and you can revoke it at any time.
Add Ayedesk to your agent

Install the plugin in Claude Code, or add your desk as a connection in Codex or your MCP client, with the token.
Ask for an aye

Tell your agent to post its work to Ayedesk before it acts. It writes the card, waits, and carries on once you decide.

Docs for developers: coming soon.

The capability contract

What an agent can do, and what it never can.

One manifest defines every operation, which tokens may call it and the rules around it. The tools, the guide and the API's permission table are all generated from it, and tests probe every route with every kind of token.

Agents can

Cards: post and resubmit after feedback.
Decisions: wait, claim once, reply to notes, recover through the inbox.
Plan: read the roadmap, save work items, comment.
Docs: search, read, write pages, answer notes, propose revisions.
Files: attach evidence to a card or a check.
Activity: report what they're working on and what it cost.

Only people can

Decide, undo or reopen a card, tick checks and write their notes.
Change a card's priority.
Accept or reject proposals and resolve doc notes.
Triage the backlog; set work items done or dropped.
Start, complete or reopen milestones.
Manage tokens, projects, access, members and workspaces.

An agent that tries a people-only action is refused, and there is no tool for it. Outside its projects, an agent can't even tell that a project exists. Card, page and comment text is treated as data, never as instructions.

Token budgets

Cheap for a model, and measured.

Every build runs scripted agent sessions against the real app and counts what each step costs. The build fails if a budget grows. A fresh session pays 2,307 tokens once, for the core tools and the guide; after that, compact responses keep each step small.

Agent token budgets: what each step costs today and its target
Step	Tokens today	Target	Status
Core definitionsDiscovering the core tools: every name, description and input	1,709	2,000	Within target
GuideReading the agent guide	598	600	Within target
Post a 3-bullet cardWriting the card and reading the answer	180	450	Within target
Wait (largest)Waiting, and the one line it reads when you decide; nothing while idle	385	200	Over target, held at 385
Answer 3 doc notes and proposeThree one-line answers to doc notes, and a proposed revision	559	1,500	Within target
Update one roadmap itemReading a 20-item roadmap and saving one work item	578	1,000	Within target
Search and read a 4,000-char docA search, and reading the page it finds	1,334	1,400	Within target

Tokens are estimated as characters ÷ 4 of what the model writes and reads. Where a step is over its target, today's value is held as a ceiling: it can only go down.

The agent guide

One page every agent reads.

598 tokens that say what an agent can do, what it never does, and the eight rules for posting, waiting and answering. Every integration gives your agent the same text.

What the guide says