← Back to Markboard

Markboard MCP server

Last updated: 2026-08-19.

Markboard is a markdown-native Kanban board. This page documents its Model Context Protocol server, which lets an AI assistant read and update your boards on your behalf — turning a plan into cards, filing notes onto a card, or answering questions about what is in progress.

Endpoint

https://markboard.work/api/mcp

Transport is streamable HTTP. You need a Markboard account; sign up free at markboard.work.

Connecting

ClaudeOpens Claude, prefilledCursorInstalls in one clickVS CodeInstalls in one clickLM StudioInstalls in one click

Claude, Cursor, VS Code and LM Studio open straight into their own add-a-server screen — Claude's arrives with the name and address already filled in for you to confirm. ChatGPT has no equivalent link, so its card copies the address and opens the settings page you paste it into. Claude Code, Codex CLI, Gemini CLI and OpenCode copy one command that adds Markboard andopens the sign-in page — OpenCode's covers its desktop app too. Either way you approve the connection here, and choose then what it can reach.

Underneath the buttons there are two ways in, and which you want depends on one question.

OAuth — for an assistant you use interactively

Add the endpoint above as a connector in Claude, ChatGPT, Gemini CLI or any MCP client that speaks OAuth. You will be sent to Markboard to sign in and approve the connection. Nothing needs to be copied or pasted, and no key is ever shown to you.

Markboard is its own OAuth 2.1 authorization server. It supports Client ID Metadata Documents and Dynamic Client Registration, PKCE (S256 only), and resource-bound access tokens with rotating refresh tokens. Discovery is at /.well-known/oauth-authorization-server and /.well-known/oauth-protected-resource.

ChatGPT: use https://markboard.work/api/mcp?client=chatgpt. ChatGPT expects authentication errors in a different shape from every other client, and that parameter is what selects it. It is the same server and the same account.

API key — for something with no browser

A cron job, a script, a local model or CI has no one to click “approve”, so OAuth does not apply. Create an API key on your profile page and send it as Authorization: Bearer <key>. The key is shown once at creation and stored only as a hash.

What a connection can reach

You choose the reach when you connect, and it is the same set of choices for an API key:

Write access always requires picking a single workspace or board; there is deliberately no “write to everything” option. A connection acts as you and never more: it cannot see a board you cannot see, and being added to a board later does not widen a connection that was scoped to something else.

What the tools do

24 tools. The read tools list boards, fetch a board or a card, search cards, and read the text of a .md or .txt attachment. The write tools create and move cards, rename them, set descriptions, dates and completion, manage checklists, labels, comments and attachments, create boards and columns, and archive or restore a card.

There are no destructive tools, by design. Nothing an assistant can call permanently deletes anything. Archiving is reversible and has its own tool to undo it; deleting a card, board, workspace or attachment for good is only possible by a person, in the app.

Agent activity is not hidden: every write is recorded in the board's activity log, badged with the name of the app that made it, and appears live to anyone else viewing the board.

Disconnecting

Every connected app and API key is listed on your profile page, with what it can reach and when it was last used. Disconnecting takes effect immediately — access and refresh tokens are destroyed at once rather than expiring on their own schedule.

Limits

Requests are rate-limited per connection, with a tighter budget for writes and for file uploads. Attachments are capped at 5 MB, and a card holds at most 5 images. Exceeding a limit returns an error, not a silent truncation.

Data and privacy

When an assistant reads a card, that card's content is sent to whoever operates the assistant and is handled under their privacy policy. Nothing leaves Markboard unless a tool is called. See our Privacy Policy for what we store, and our Terms of Service.

Support

Questions, a bug, or a security report? Email support@markboard.work.