A research desk for every chess broadcast

Streamer64 puts a purpose-built chess database at a commentator's fingertips while the game is live: what masters do in this structure, how these two players handle this opening, what changed in modern chess. It is built for human commentators. The AI booth on this site is a demonstration of what it can supply.

For streamers and commentators

You bring the voice; Streamer64 brings the research. While you talk, it reads every position against the millions of games in f64db and puts what matters in front of you, a click away from your viewers' screen:

the head-to-head, and each player's best and worst openings,
how masters, engines and correspondence players handle the opening, and the move orders that reach it,
the plans in the pawn structure, with model games you can play through on a demo board,
how often an endgame like this one is held, and how top players did it.

A word of honesty: we know a little chess and even less about streaming. What we know is f64db and what it can do. So if you want to start streaming with it tomorrow, we're not ready for you yet: as of now, Streamer64 is a preview release.

Questions about the database

Do I have to use the TWIC games?: No. Bring your own games as PGN and f64db will work with them.
What about licensing?: Get in touch, and we'll talk it through.
Do I need an AI or LLM subscription?: No. Streamer64 gives you the data points for each position straight from f64db. The AI commentary on this site is only a demonstration.
Do I need servers?: No. Streamer64 runs on a Windows PC, and Mac support is on its way. You'd only need a server to open a web page like this one to the public.
Do I need an internet connection?: Only to follow live games. Recording highlights of games and publishing them later works offline.
How many games can it handle?: Over 15 million games on a good machine, without breaking a sweat.

[email protected] Watch a recorded game first

What it gives a commentator

Every move, Streamer64 asks f64db about the new position, and the answers arrive in milliseconds from millions of games.

What strong players did hereThe moves and whole lines played from this exact position in strong classical games, with results and an example game.

The pawn structure's plansEvery game with the same pawn structure, and how masters changed it: the typical breaks and how they scored.

The players themselvesEach player's record in this opening by colour, the times they reached this very position, the head-to-head, and where their own experience ends.

Strengths and weaknessesOpenings and types of position where a player scores well above or below their own average, with the games behind each figure.

Move ordersThe different orders by which recent games reached this position, how often each was taken, and the options each order gives the other side on the way, drawn as a map.

Novelty alertsMoves first played in this position in the last three months, who played them first and how often since, and an alert when a player has just played one.

FashionWhat the elite play here now against all-time, and which moves are rising or fading.

Miniatures and endgamesShort decisive games from this position or structure, and endings with the same pawns, material and kings, with how top players handled them.

The numbers on the boardSpace, activity, the centre, pressure on each king, loose pieces, forks and pins, counted from the board, with no engine involved.

How a player thinks hereThe opponent's threats first, then checks, captures and threats, then the candidate moves, and whether the position asks for deep calculation or a choice between plans.

Endgames by their own rulesIn the endgame the talk turns to king activity, passed pawns, the opposition and where the rooks stand, not to king safety.

The engine, second and labelledStockfish is consulted after the human discussion, and always named as the engine. A forced mate is the exception: it is confirmed at once.

How a commentator uses it

Indicators. Chips light up as the database finds something worth discussing in the position: a player has been here, the structure has known plans, a miniature exists. Click one to put its graphic on screen.
Research tools for any assistant. The same database tools are offered to AI assistants through the Model Context Protocol, so a commentator's assistant can research a position and put a line or a graphic on the broadcast screen.
Show. A demo board plays the line being discussed, labelled DEMO, beside the live board. Stat graphics and pull-up cards (positional, tactics, the numbers, how a player thinks, move orders, the scoreboard) come up on demand or with a key.
Engage the audience. At a tactic or a forced mate, viewers get a timed "your move" challenge on the board. Their picks are aggregated and judged, and when the player moves the booth compares the player's choice with the viewers'.

The director

A broadcast needs a producer as much as a voice. The director decides when something is said and what comes next:

A forced mate, or a mate threat, is discussed before anything else.
Then the position is read, the plans are discussed with the data, and the players and history come in during long thinks.
A topic is said once per game, and talk about an old position is dropped the moment a new move lands.
Recaps, clock remarks and supporter messages keep their own schedules. A viewer joining mid-game is brought up to speed.
The broadcaster chooses which games go on air, up to five at once. A game on air with nobody watching costs nothing: no commentary is written for an empty room.

How it fits together

Lichess broadcast→director→f64db · board numbers · engine (gated)→commentator or AI booth→voice · boards · graphics

Games: live Lichess broadcasts, one shared stream per round however many games are followed.
Data: f64db, opened read-only. The same database tools serve the booth, the copilot, and any agent through an MCP server.
Output: a web page with live and demo boards, clocks, captions, graphics and audio, ready to capture into a stream.
Grounding: every move and number said is checked against what the tools returned, and a line that fails is dropped, and every number is said with its source. When a figure is counted from the board, its panel is on screen.

About the AI commentary on this site

It is a demonstration, with a human in the loop as the goal. The commentary uses a large language model in its most rudimentary form, with no special training and no commentary skills given to it, so it can make mistakes. It could be improved, but that is not our objective. Our objective is to show what Streamer64 and f64db can put in the hands of a human commentator.