The AI assistant and the MCP server

What the assistant does

The AI assistant, at /en/dashboard/ai-assistant, is a chat over the workspace you are working in. It answers from that workspace's own data in the database: its projects and their boards, the tasks by column, priority, person or due date, the members, roles and pending invitations, the plan's usage, the billing and its invoices, the launch checklist and the recent activity.

• Answers stream as they are written, in the language the dashboard is read in, with cards for lists, tables, usage and the checklist, and charts of task activity, API requests or projects created, per day or per month. Each reply ends with three follow-up questions.
• The empty screen offers openers worked out from the workspace itself, such as how many projects need attention or how many invitations are pending, with no model involved.
• Up to six images can be attached to a question, 10 MB each.
• It never changes the workspace by itself. When asked to plan work, it prepares tasks as a card that a person applies or dismisses.

Tasks it prepares

Asked, for example, to "Draft a launch checklist for Website relaunch", the assistant answers with a card such as "Add 3 tasks to Website relaunch", listing each task and its priority.

• Add 3 tasks puts them at the end of the project's backlog, once, and the card offers "Open the Website relaunch board". The access is checked again at that moment, so a role changed since the card was made is respected.
• Dismiss adds nothing: "Dismissed. Nothing was added."
• A member who may read the card but not add tasks sees "Your role can read this suggestion but not apply it."

The seeded conversation "Launch readiness" holds one such card, waiting to be applied.

Conversations

• Every conversation is saved with its messages, so it survives a reload. New conversation starts another, and each one can be renamed, archived or deleted from its menu under Conversations.
• Each member sees their own conversations only, in the workspace they were held in.
• Forma Studio comes with three sample conversations, "Launch readiness", "Usage summary" and "Team access", which every member can read but nobody can change. A question asked in one of them starts a conversation of the reader's own.

What it needs, costs and who can use it

• An account and a key with Google Gemini, OpenAI or DeepSeek. Any one is enough. The server's keys go in the API's .env as GOOGLE_API_KEY, OPENAI_API_KEY and DEEPSEEK_API_KEY.
• Without a server key for the chosen model's provider, the assistant still works: sending a message opens Use your own API key, where each person can add a key of their own. It stays in their browser and is sent only with their own requests.
• Each answer is billed by the provider of the model picked, at that provider's price, to the key that paid for it: the person's own key when they added one, otherwise the server's. The template itself charges nothing.
• The ai.read permission to open the assistant. Of the seeded roles the Owner, the Admin and the Member hold it; the Billing role does not, and the assistant is not in its sidebar.
• Nothing beyond the member's own role: the assistant reads only what that member's permissions allow in that workspace, and checks again on every step. Applying a card of tasks needs ai.apply and projects.write.
• The API. On the built-in sample data the assistant answers that it needs one.

• AI keys in the configuration
• When the assistant asks for a key

The MCP server

The API also offers the assistant's tools over the Model Context Protocol, so a coding agent such as Claude Code, or any MCP client, can read one workspace's projects, tasks, team, usage, billing and activity directly, and prepare tasks for it. It needs a key you choose, and refuses every request until one is set.

Models

Whose key pays

The tools

Connecting an MCP client