MCP: your own AI, inside Flowra

What is MCP?

MCP (Model Context Protocol) is a common language AI assistants use to connect to apps. Flowra has an MCP server: your AI sees Flowra's list of tools (like «create action» or «search Docs») and calls them when you ask it to do something.

The thinking is your own AI's job and Flowra doesn't charge for it. Each tool goes through the app's own paths, so access, notifications and action history follow the usual rules.

Connecting

There are two ways. Claude Desktop, Cursor and VS Code connect with a personal token that is shown once and kept like a password. claude.ai and ChatGPT connect through OAuth and bring you to Flowra's approval page; this only works when Flowra is on a public https address.

1. Open “MCP and AI agents” from the Connectors hub
2. Press “Create token” and choose permissions and expiry
3. Copy the token right away and put it in your AI app's settings
4. Revoke the token whenever you want

What Flowra exposes

Flowra's MCP server has twenty-nine tools in ten sections, and your AI sees only the tools its token allows. General: “who is this connection acting for” and search across all permitted sections. Workspaces and projects: the list of workspaces and your role, the list of projects, and a single project with its open actions and Docs pages. Actions: a filtered list, reading one action with its checklist, comments and history, creating, changing status, priority, due date and assignee, commenting with mentions, and working with the checklist.

Agent queue: see the queue, pick up, report progress and finish. Docs: search, read a page as Markdown, create a page (linked to an action or project) and append to the end of a page. Meetings: list, read a meeting with its summary, decisions and transcript, and create a meeting with a conflict check. Notes: read and create. My Network: list and read contacts, log an interaction, list opportunities and move their stage. Chat: list, read and send messages in your name. Finance: only a read-only summary of balances and this month's income and expenses.

Alongside the tools there are four ready-made prompts: “Work on this action”, “Work through the agent queue”, “Plan this project” and “Summarise this meeting”. Projects, actions, Docs pages and the agent queue are also available as readable resources. My Files isn't available through MCP yet; coming soon.

Creating and revoking tokens

Give each app its own token so you can cut each one off separately. A token starts with flw_mcp_, is shown only that one time, and Flowra keeps only its hash; if you lose it, revoke it and create a new one. Each account can have at most twenty active tokens.

In the list, each token shows its name and agent, its first four characters, “Personal token” or “Via OAuth”, its number of permissions, last use and expiry. “Revoke” asks “Revoke this token?”, and any app connected with it is cut off that very moment; this can't be undone. You can't change a token's permissions from this page for now; create a new token with the right permissions and revoke the old one.

1. In “Tokens”, write a name, like Claude laptop
2. Turn permissions on or off; at least one is required
3. Choose the expiry: no expiry, 30 days, 90 days or one year
4. If the token is for an agent, choose it under “On behalf of which agent?”
5. Press “Create token” and copy the token right away

Connecting Claude Desktop, Claude Code, Cursor and VS Code

Under “Server address and setup”, each app has a tab with steps and a sample config. The server address is always your Flowra address plus /api/mcp, and the token is sent in the Authorization header as Bearer, not in the address. Once you create a token, <YOUR_FLOWRA_TOKEN> is filled into every sample automatically.

Claude Code connects with one command: claude mcp add --transport http flowra https://<your Flowra address>/api/mcp --header "Authorization: Bearer <YOUR_FLOWRA_TOKEN>". Cursor takes an mcp.json file with a flowra entry under mcpServers that has url and headers, and VS Code the same in .vscode/mcp.json with type set to http.

Claude Desktop only runs local servers, so the claude_desktop_config.json sample uses npx and the mcp-remote package to bridge to Flowra's server, reading the token from the FLOWRA_AUTH variable. After saving the file, quit Claude Desktop once and open it again. There's also a curl sample for testing by hand.

claude.ai and ChatGPT with OAuth

claude.ai and ChatGPT don't copy a token. Add the server address in that app's connectors; the app brings you to Flowra's “Connect an AI app” page, where you see the app's name and the permissions it asks for, can switch any of them off and choose an agent, and answer with “Allow” or “No”.

A token created this way is valid for thirty days, and the app renews it with a one-hundred-eighty-day refresh token; a renewal never gains more permissions. This only works when Flowra runs on a public https address. An OAuth connection also appears in the token list as “Via OAuth” and is revoked right there.

Permissions only restrict

Each token has only the permissions you chose, such as reading or writing actions, knowledge, calendar, notes, contacts and chat. Finance is read-only and has no write permission at all. A token can never do more than you: the AI sees only what you see.

Every call is logged under your name, and your recent calls are visible on this page. An admin account can't create tokens.

Agents and the work queue

An agent is an identity you can hand actions to, such as “Claude”. Press “Send to Claude” on any action's page; your own Claude, connected with this agent's token, picks it up, comments on its progress and finishes it with a summary. Every step is written to the action's history and you are notified.

An agent can be personal or belong to a workspace so members can send it work too. An action only joins the queue if the agent's owner can edit it themselves.

Handing off an action from its page

On every action's page there's a “Hand off to an AI agent” card. If you have more than one agent, pick one, write the “Note for the agent” and press “Send to …”. The status badge right there tells you whether it's “In the agent's queue”, the “Agent is working” or the “Agent finished”, with the “Latest report”. While the work is open, there's also “Cancel hand-off”.

You can send only when you can edit that action yourself and the agent's owner can too. Sending an action that's still open again returns the same earlier job. If someone else sends work to your agent, you get a notification.

Examples: how Claude works

These are things done with these very tools today. Just ask Claude in your own words; it picks the right tool.

1. “Work through the agent queue”: Claude reads the queue and picks up the first action; its status turns to in progress
2. Midway it reports its progress; each report is a comment in the action's history
3. “Move this action to review and set its due date to Friday”: one change on that same action
4. “Write the research summary in Docs”: a new Docs page, linked to that same action
5. At the end it closes the work with a summary and you get a notification

Security and audit

Every request passes these checks in order: the request's origin, the token, the MCP switch in the admin panel, the rate limit, and for each tool the token's permissions and then Flowra's own rules. A tool without the needed permission doesn't run, and it tells your AI which permission is missing.

Every tool call, every denied request and every token creation or revocation is recorded in the audit log, with the app name, time and result. “Recent calls” at the bottom of this page shows the latest things your AI did in your name, or asked for and was denied, with the reason: the token's permissions weren't enough, Flowra didn't allow it, or the input was invalid.

Comments and pages written through MCP are signed with your name and the agent's. The super admin account and deactivated accounts can't have MCP.

Constraints

Each token gets six hundred requests every fifteen minutes and then waits fifteen minutes; more than thirty wrong tokens from one address block that address for fifteen minutes. Each account has twenty active tokens and ten agents.

The server is stateless and works over Streamable HTTP; there's no local server or SSE stream. Finance is read only, and no tool moves money. My Files, editing a token's permissions and editing an agent from this page aren't here yet; coming soon.

On phones

The MCP page is complete on your phone too: create, copy and revoke tokens, and create agents. The hand-off card is on the action page on your phone as well, so you can send work to Claude's queue from anywhere.