
pokeragile
io.github.florianmousseau/pokeragile
Planning poker rooms an AI assistant can join and vote in, its card hidden until the team reveals.
Documentation
PokerAgile
Planning poker for agile teams, in the browser. Create a room, share its link, vote, reveal. No account, no install.
Live at pokeragile.com, in English, French, Spanish, German and Portuguese.

What it does
- Rooms by link: the home page creates a room, and anyone with the URL can join by typing a name.
- Card decks: Fibonacci, T-shirt and powers of two presets, or any comma-separated list.
- Real-time votes over a WebSocket. Votes stay hidden until someone reveals them. The results then show the average, median, min, max and distribution.
- Observers can follow a room without voting. Anyone in the room can switch a participant to observer or remove them. Removing someone asks for confirmation first.
- Host and lock: the first person in a room is its host and can lock it, so that only they reveal, start a new round or remove people. Off by default.
- Voting timer: 30 s, 1, 2 or 5 minutes, shown to everyone.
- Story list: paste stories one per line, estimate them in turn, keep the chosen card for each, export as CSV or a Markdown table.
- Agreement and auto-reveal: the share of votes on the most chosen card, the lowest and highest votes marked, and an optional reveal on the last vote.
- AI participants: an assistant joins a room over MCP and votes like anyone, marked as an AI. See below.
- Reconnection: a dropped connection retries with exponential backoff. A reload within 30 seconds takes the same seat back, vote included.
- Ephemeral: rooms live in the server's memory and are deleted after 24 hours without activity. There is no database.
- Accessibility: semantic HTML, full keyboard use, and live regions that announce votes and results. The site respects
prefers-reduced-motion. It targets WCAG 2.1 AA; the accessibility statement currently reads "partially conformant".
Let an AI assistant vote
The room server speaks the Model Context Protocol. Add this address as a remote MCP server (a custom connector in Claude, an MCP server in ChatGPT, Cursor or VS Code):
https://salles.pokeragile.com/mcp
Then ask the assistant to join the room the team shares and estimate the story. Its tools:
| Tool | What it does |
|---|---|
join_room | Takes a seat marked as an AI and returns a seat token. An AI never holds the host role. |
get_room | Reads the deck, the round, the story list and the one on the table, who voted, and the results once a person reveals them. |
add_stories | Appends story titles to the room's list, 100 stories of 200 characters at most. Refused when the host locked the room, as for any participant. |
vote | Votes one card of the deck with a short reason, both hidden until the reveal. |
get_estimates | Reads the story list with the estimate the team kept for each, as data and as a Markdown table. |
leave_room | Gives the seat up. |
An assistant that already has access to Jira, Linear or GitHub can do the whole tour: read the sprint's stories there, add them to the room, vote on each, then read the kept estimates and write them back. PokerAgile connects to no tracker; the assistant uses its own access.
Take the stories of the current sprint in Linear and add them to my room https://pokeragile.com/ab12cd34.
Vote on each one when it is on the table, with a one or two sentence reason.
When I say we are done, write the estimates the team kept back to Linear.
An assistant cannot reveal, start a round, choose the story on the table, keep an estimate, clear the list or remove anyone, and it sees nobody's vote before the reveal. Its reason shows under its name at the reveal. Nothing changes in a room nobody connects an assistant to. No account and no key: the endpoint is public, like the rooms. PokerAgile calls no AI service; the assistant reads the room because its user connected it.
The browser's built-in AI
Where Chrome exposes its on-device model (the Prompt API, Chrome 148 and later on desktop), the room shows a button that analyzes the story on the table: what it leaves out, and a split if it looks too large. The model runs on the reader's computer; the analysis sends the story to no service and the answer is shown to that reader only. Elsewhere the button does not appear.
Privacy
No analytics script, no advertising cookie, no visitor identifier, nothing loaded from a third party. Like any website, the host that serves the pages counts the requests it answers. The site also counts its own readers, server-side: a daily fingerprint derived from the IP address, which is never kept. Both counts are server-side and never touch the reader's device. npm run mesure checks that the pages keep saying both halves of this.
Run it locally
Requires Go (the version in backend/go.mod) and Node 22.
./dev.sh
Then open http://localhost:5173. The script creates frontend/.env from .env.example and installs the frontend dependencies on the first run. It then starts the backend on :8080 and the Vite dev server on :5173. Ctrl+C stops both.
On Windows without a POSIX shell, run the two halves in separate terminals:
cd backend && go run ./cmd/server
cd frontend && npm ci && cp .env.example .env && npm run dev
Architecture
| Part | Runs on | Talks to the browser over |
|---|---|---|
| SvelteKit 5 frontend: pages, SSR, security headers | Cloudflare Pages | HTTP |
| Go backend: rooms in memory, one goroutine per room | OVHcloud VPS, Gravelines (France), built from main by the server itself | WebSocket, plus POST /rooms |
backend/is Go with the standard library HTTP server andgolang.org/x/net/websocket.internal/storeholds the rooms in a map.internal/roomowns each room's state behind one event loop, which broadcasts a snapshot on every change and masks the votes until they are revealed.internal/handlerexposesPOST /rooms,GET /rooms/{id}/ws,/healthand/stats.frontend/is SvelteKit 5 with runes, TypeScript and plain CSS, deployed withadapter-cloudflare.src/lib/useRoom.svelte.tsholds the WebSocket client andsrc/routes/[id]the room page. The room translations live insrc/lib/i18n/, one file per language, fetched one language at a time bysrc/lib/dictionaries.ts.- Security headers are set in two places that must stay in sync.
frontend/src/hooks.server.tscovers pages rendered by the worker andfrontend/_headerscovers static files.
Checks
Every pull request runs the same steps as the Code Quality and Bundle Size jobs of deploy.yml:
| Command | What it checks |
|---|---|
golangci-lint run (in backend/) | errcheck, govet, staticcheck, ineffassign, unused, bodyclose |
go test ./... | backend unit and WebSocket tests |
npm run lint | ESLint on TypeScript and Svelte |
npm run check | svelte-check types |
npm test | Vitest, including French typography rules |
npm run check:seo | titles, descriptions and alternates of every page |
npm run check:reflow | CSS that could push a page past a 320 px screen |
npm run mesure | no measurement script, and the privacy wording above in five languages |
npm run build && npm run size | a first visit downloads under 50 kB of JavaScript (brotli), on the home and in a room |
CodeQL scans Go and TypeScript, and SonarCloud analyses the code on each run. After each production deploy, Lighthouse CI audits the live site.
Contributing
See CONTRIBUTING.md.
License
No installation packages available.