Aback Chess API and MCP server
A public REST API with no key, an MCP server for chess tools, and OAuth 2.1 access to a player's own account for AI agents.
On this page
Quickstart
The public API needs no key and no account: send a GET and read the JSON. It is read-only and every answer is the same for everyone. There is no sandbox because nothing here can be changed; call the live endpoints.
curl https://abackchess.com/api/v1/puzzles/daily{
"ok": true,
"data": {
"day": "2026-11-13",
"puzzle": { "id": "…", "fen": "…", "moves": ["e2e4", "…"], "rating": 1650, "themes": ["fork"] }
}
}Endpoints
- GET /api/v1/stats: the site's totals, all time: players signed up, games reviewed, puzzles solved and games played against the computer. Refreshed every few minutes.
- GET /api/v1/puzzles/daily: the puzzle of the day for the UTC day, from the Lichess puzzle database (CC0). moves is in UCI: the first move is the opponent's, the rest is the solution. 404 with NO_PUZZLE if none is set.
- GET /api: the API's front door, listing these endpoints and the documents below.
The OpenAPI 3.1 description is /openapi.json, with typed schemas for every answer and error, ready for function calling. The same docs in Markdown are /docs/api.md, and the RFC 9727 catalog is /.well-known/api-catalog.
Errors
Every answer uses one envelope. A failure has a machine-readable code to branch on, a message for a person, a hint on what to do, and a link to these docs. Any address under /api that does not exist answers 404 in the same JSON, never an HTML page.
{
"ok": false,
"error": {
"code": "RATE_LIMITED",
"message": "Too many requests from this address.",
"hint": "Wait 12 seconds, then poll less often.",
"docs": "https://abackchess.com/developers"
}
}Codes: NOT_FOUND (404), NO_PUZZLE (404), RATE_LIMITED (429), UNKNOWN_ERROR (500) and DATABASE_UNAVAILABLE (503).
Rate limits
About 120 requests a minute from one address. Every answer carries the IETF RateLimit and RateLimit-Policy headers, and RateLimit-Limit, RateLimit-Remaining and RateLimit-Reset for clients that read those. Past the limit you get a 429 with Retry-After in seconds. The answers change every few minutes at most, so once a minute is plenty.
Versioning and deprecation
This is version 1, at /api/v1/, and every answer says so in its API-Version header. A change that could break a client gets a new path (/api/v2/). The old version keeps answering for at least six months after the new one ships, and during that time its answers carry the Deprecation and Sunset headers (RFC 9745, RFC 8594) naming the date it stops, and a Link header with rel="deprecation" pointing here. No version has been deprecated yet. Adding a field is not a breaking change, so ignore fields you do not know. The unversioned paths /api/stats and /api/puzzles/daily answer as version 1.
MCP server for chess tools
AI agents can use the same answers and more over the Model Context Protocol at /mcp (Streamable HTTP, no key, read-only): the daily puzzle and site totals, a search of the chess guides and glossary, a check of any FEN, reading a PGN with its opening, and a link that opens a game in the Game Review. Its server card is /.well-known/mcp/server-card.json, and every tool is described on the MCP server page.
claude mcp add --transport http aback-chess https://abackchess.com/mcpA player's own account, for agents
With a player's permission, an agent can read their saved games, Game Reviews, puzzle rating and practice progress through a second MCP server, /mcp/account. Access is read-only and scoped:
- games:read: Read the games saved to your account, and their Game Reviews
- profile:read: Read your puzzle rating and your practice progress
There are no API keys to request. An MCP client connects with OAuth 2.1 as the MCP authorization spec describes: it registers itself at /oauth/register (open dynamic registration, no secret), the player approves it on our consent page, and PKCE and rotating refresh tokens do the rest. Discovery starts at /.well-known/oauth-protected-resource and /.well-known/oauth-authorization-server.
An agent that knows only its player's email address can register by email instead, as /auth.md describes: it asks at /agent/identity, hands the player a six-digit code and a link, and collects its token once the player has typed the code here. Players see every connected agent on their account page and can disconnect it at any time.
What is not offered
There is no analysis API: the chess engine runs in the visitor's browser, on their own device. To have a game reviewed, send the visitor to the review with the game in the address, as the agent skill at SKILL.md explains. Everything under /api other than the endpoints above belongs to the site's own pages and may change without notice.
Support
Questions about the API or the MCP servers go to support@abackchess.com. The terms of use are on the terms page, and a summary of everything for language models is /llms.txt.