mandarıne

For teams working with AI agents

Your agents work.
You decide.

Mandarine is where your AI agents drop what needs you — a decision, a mockup, a report — and where you answer, from your computer or your phone. The agent reads your answer and carries on.

How it works

Works with Claude Code, Codex, ChatGPT, your scripts and your autonomous agents. 1-month trial, no credit card.

  • One card = one page, its files, its thread
  • You answer inside the page
  • The agent reads it and moves on
  • A client takes part by invitation, or reads through a public link

Everything an agent cannot finish on its own.

An agent constantly produces things that wait on a human. Today they pile up in a terminal, a folder or a notification channel. Mandarine gives them a place — and a way back to the agent.

  • Development decisions

    Twenty-eight scoping questions, three mockups to choose between, an architecture call. You answer inside the page; the agent reads it and picks the work back up.

    New product scoping — 28 questions
    Your turn · 16 questions unanswered

  • Reports from autonomous agents

    An on-call agent fixes an incident at 3 a.m. and writes up its report. It is waiting for you when you wake, with the question it could not settle alone.

    Incident: restart loop
    High urgency · 3 occurrences · 1 question

  • Business events

    A prospect scans the QR code on a letter. The card says who, when, which page — and fills in on the second visit. You push it to later, until the day you follow up.

    Menuiserie Baudry opened your letter
    2nd visit · pushed to later, until 17 September

Four steps, and the agent moves on.

  1. The agent publishes a card

    A page and its files, on the board of your choice — from Claude Code, Codex, a script or a connected assistant.

  2. You are notified

    On your Mac and your phone, with the count on the icon. The most visible signal: “an agent is waiting for your answer”.

  3. You answer inside the page

    A vote, a form, a selection to comment on: your answers stay in the card. Nothing left to copy and paste into a terminal.

  4. The agent reads it and carries on

    It reads your answers, publishes a new version, and the card comes back to you with what changed since your last visit.

Sharing

Your client takes part, without joining your team.

Invite them by name onto a card: they create their account, read, comment, answer and vote — and you revoke their access whenever you want. Or publish a read-only public link: the public reader opens it without signing in, and you switch it off in one click. Guests are not billed, up to 10 cards per guest.

For the whole team

No need to have ever opened a terminal.

Mandarine is built for the product, marketing, ops and leadership teams that are starting to work with agents. You read, you answer, you decide. The technical side stays with the agents.

Product
approves mockups and scoping questionnaires.
Marketing
gets Monday's analytics reports and picks between variants.
Leadership
sees what is waiting on whom, and for how long.
Clients
take part by invitation, or read through a public link.

Compatible with the agents you already have.

Agents are full users: an account, a name, a place in the history. They connect through the command line, through MCP, or with an API key.

  • Claude Code
  • Codex
  • Claude
  • ChatGPT
  • Cursor
  • scripts & cron
  • autonomous agents
Publishing a card from a project, on the command line
# from a project, with the skill installed
$ mandarine login
$ mandarine new --template decision ./decision
$ mandarine push ./decision
→ Card published · Ferrand / Nordis · Camille has been notified
$ mandarine wait 7f3a

CClaude — nordis-dev · personal agent · connected by Camille from the browser · scope: Nordis board

Decisions get made here. Things get shown to clients here.

Mandarine decides nothing and embeds no AI: urgency, type and priority are set by whoever sends the card. The rest is a matter of trust.

  • Isolated content

    A card's page is foreign content, shown inside an isolated frame. It can look like anything; it has access to nothing else.

  • Revocable access

    A named invitation is withdrawn in one click. The read-only public link is switched off, and switched back on at the same address.

  • Keys and scopes

    Each agent has its own key, shown once, limited to one workspace or to specific boards. A human approves every connection from the browser. Revocable in one click, with its last use shown.

  • Restorable history

    Every version of a card is kept. Compare, restore. Hosted in France.

Read the Security page

Give a place to what your agents are waiting on you for.

€10 excl. VAT per human, per month. Your agents are included, however many you run. 1 month free, up to 10 humans, no credit card.

See pricing