Agents

Agents can write. Only people publish.

Connect Claude or another assistant to a project, and it follows the same rules as your team. Its changes carry its name and stay marked until someone reads them. Going live takes a person.

In a terminal, Claude Code updates an article’s excerpt through Moteur’s MCP server and asks for a review. Then Céleste opens the review: “Louise Gagnon via Claude Code asks for a review”, the new excerpt marked as written by Claude Code and not yet read. She reads it, and approves.

Claude Code asks for a review. Céleste reads what it wrote, then approves.

Try it

Choose what it may do

Louise connects Claude Code to the Moteur Journal. Pick the access she gives it, and see what each of its requests comes to.

What it asks, and what happens

  1. Read the five newest articles

    Goes throughRead. Nothing changes.

  2. Shorten an article’s excerpt

    Goes throughSaved, and marked “Written by Claude Code · not yet read” until someone reads it.

  3. Ask Céleste to review it

    Goes throughCéleste sees “Louise Gagnon via Claude Code asks you to review”.

  4. Publish it

    RefusedChanges made by an agent go live once a person approves them: ask for a review.

  5. Approve its own review

    RefusedAn agent can’t approve or reject reviews: a person does.

  6. Add a “Listening time” field to Article

    RefusedThis tool is outside the connection’s current permissions

  7. Delete an old draft

    RefusedNot offered: tools that can’t be undone are left out unless the connection allows them.

The levels and the consent screen are the Studio’s. The refusals are the engine’s, word for word. The workflow here has reviews on.

The rules

Where agents fit

Agents work inside the project’s rules, not around them. Your client always knows what an assistant wrote, and nobody has to trust it blindly.

  • Names

    Two names on every change

    An agent always works for a person: “Louise Gagnon via Claude Code” in reviews, history and the project’s Git log.

    • Never anonymous
    • The person’s own role still applies
  • Marks

    Marked until read

    Every field an agent writes is marked “not yet read”. It can’t clear its own mark.

    • “Looks right” clears it
    • Shown in reviews too
  • Reviews

    Review-only

    An agent can ask for a review, never approve one. Connected through Moteur Cloud, it can’t publish at all.

    • It can’t loosen its own rules
    • A person approves, every time
  • Connections

    Connections you can see

    Every assistant you’ve connected is listed under AI connections, with its access and projects. Disconnect one, and it stops at once.

    • Projects picked one by one
    • Each connection lasts 30 days
  • Safety

    Nothing it can’t undo

    Tools that delete, purge or merge are left out unless you allow them. Every save is a commit, so any version can come back.

    • Off unless you say so
    • A way back from every save
  • Clients

    The assistant you already use

    Claude Code, Claude, ChatGPT and other MCP clients connect to Moteur’s hosted server, or to a local one signed in like the CLI.

    • Set up in a minute
    • The same rules either way

Questions about agents

An agency owner asksCan an agent publish on its own?

MoteurNot through Moteur Cloud: its changes wait for a person’s approval. A developer running one on their own machine can let it publish their own work.

A reviewer asksWill I know what the agent wrote?

MoteurYes. Every field it changes is marked until a person reads it, and every change carries both names.

A client’s editor asksCan it see all my projects?

MoteurOnly the ones you pick when you connect it. Your other projects, and future ones, stay private.

An agency owner asksAnd if it goes wrong?

MoteurDisconnect it under AI connections, and its access stops at once. Every save it made is a commit, so any earlier version can come back.

Let an agent do the first draft.

Start a free project, connect Claude, and read what it wrote before anyone else does.