I maintain PokerAgile, a free planning poker app: one person opens a room, shares the link, everyone picks a card in secret, and the cards are revealed together. No account, MIT-licensed, Go on the server and SvelteKit in the browser.
This month I added one thing: the room server speaks the Model Context Protocol. An assistant such as Claude, ChatGPT or Cursor can join the room and vote. This post shows how it is built and what I got wrong on the first try.
What the assistant can do
The endpoint is https://salles.pokeragile.com/mcp. It is a remote server over streamable HTTP, so there is nothing to install and no API key. It exposes six tools:
| Tool | What it does |
|---|---|
join_room | Takes a room address and a name, returns a seat token |
get_room | Returns the deck, the story on the table, who has voted |
vote | Plays one card from the deck, with a one-sentence reason |
add_stories | Appends story titles to the room's list |
get_estimates | Returns each story with the estimate the team kept, as data and as a Markdown table |
leave_room | Frees the seat |
In the room, the assistant shows up as a participant marked "AI". It can never become the host, and it cannot reveal, start a round or remove anyone.
The rule that matters: the card stays hidden
Planning poker exists to stop the first number spoken from pulling everyone else toward it. An AI that announces "13" in the chat before the humans vote breaks the technique.
So the server treats the assistant like any other player. Its card and its reason are masked in every snapshot until the round is revealed. get_room hides them too, so a second assistant cannot read the first one's vote. After the reveal, the reason appears next to the card.
Three things I got wrong
1. A vote for a card that is not in the deck returned success. The assistant believed it had voted. The tool now returns a real tool error (isError) with a message the model can act on: "this card is not in the room deck".
2. The reveal waited for everyone. The browser disabled "Reveal" until every participant had voted. One assistant that joined and never voted blocked the whole team. Reveal is now available from the first vote, and the screen says "3 of 5 have voted".
3. The SDK refused the public host. The official Go SDK protects local servers against DNS rebinding and rejected requests arriving through a tunnel with a public Host header. For a server that is meant to be public, that protection has to be turned off explicitly.
From the tracker to the room and back
The two newest tools exist so that PokerAgile needs no Jira or Linear integration. The assistant already has access to the tracker on its side. It reads the sprint, calls add_stories, votes on each story when it is on the table, and at the end calls get_estimates and writes the numbers back itself.
The assistant has the rights of a participant and nothing more. When the host locks the room, add_stories returns an error that says so. No tool lets it choose the team's final estimate: a person keeps the card after the reveal, and the assistant only reads it.
Stateless on purpose
The MCP handler keeps no session. Each call carries the room and the seat token that join_room returned. Rooms live in memory, are never written to disk, and are deleted 24 hours after the last activity. The server calls no AI service: the model runs on the user's side, with the user's own subscription.
Try it
Open a room on pokeragile.com, then tell your assistant:
Join my planning poker room https://pokeragile.com/your-room through the MCP server https://salles.pokeragile.com/mcp. Read the story on the table, vote with one card from the deck and give your reason in one sentence. Do not reveal your card in the chat.
The "Invite your AI" button in the room's Options panel copies this prompt with your room address filled in.
The code is on GitHub; the MCP part is in backend/internal/assistant.
One open question I would like opinions on: when the AI says 13 and the team says 3, does that gap start a useful discussion, or does it only add noise?
