About Bot Check
Bot Check reads a StarCraft: Brood War replay and looks for habits that programs have and people don't. It runs in your browser, and the replay never leaves your computer.
Why it exists
Brood War bots have been around for as long as the game has had a programming interface, but until recently they played in their own tournaments. That changed in 2026. Pluto, a bot by tscmoo, won the CoG StarCraft AI Competition and was released to the public. Members of the Brood War AI community have reported ladder games played by an unauthorised adaptation of it, and replays from other ladder accounts show the same habits.
Watching a replay and deciding whether a player is a bot is hard. A strong bot does not move like a cartoon robot. But the way a program sends its orders is different from the way a person does, and those differences are written into the replay file. Bot Check reads them.
The three signs
Each sign comes from how Pluto actually plays. How Pluto Plays explains its design in full.
A fixed rhythm
Pluto makes one decision every 6 game frames, about four times a second. Its actions therefore land on a grid: almost every pause between two actions is 6, 12, 18 or 24 frames long. A person acts whenever they click, so their pauses come in all lengths. Bot Check looks for the coarsest grid that at least 70% of a player's pauses fit, and flags grids of 4 frames or more that the opponent does not share. Network settings can put both players on a coarse grid, which is why the opponent is compared.
Bursts
When Pluto orders twenty units to attack, it does not box them. It selects the first unit and orders it, selects the second and orders it, and so on, all within a single frame, which lasts 1/24 of a second on Fastest. A replay records each of those selections. People manage two or three selections in a frame at most. Bot Check counts frames with four or more selections and flags a player who has five or more such frames.
One unit at a time
Pluto practically never uses control groups or selects more than one unit at once. People box-select and press hotkeys constantly; in the ladder games checked, hotkeys were often a third or more of everything a player did. Bot Check flags a player when at least 99% of 300 or more selections are of a single unit and no hotkey was ever pressed.
One clear sign is enough for "Likely a bot". A sign that is present but weak, such as a single burst, gives "Unclear".
What it was tested on
| Replays | Players | Result |
|---|---|---|
| Every Pluto game from the CoG 2026 competition | 2,889 games | All flagged. All showed the single-unit selection style; 2,885 showed the 6-frame rhythm. |
| Online games between people | 40 players in 21 games | None flagged. No one made more than 3 selections in a frame, and everyone used hotkeys. |
| Ladder games with suspicious accounts | 2 accounts | Both flagged, with the same pattern as Pluto: only single-unit selections, no hotkeys, and the same bursts. |
That is a small test. The thresholds were chosen to separate these replays, and they may change as more replays are checked. Accounts are not named here on purpose.
What a result cannot tell you
- It is built to catch Pluto. The three signs come from how Pluto plays, and nearly all the bot replays it was tested on are Pluto's. Other bots, or versions of Pluto changed to hide these habits, may show no sign at all.
- "Likely a bot" is evidence, not proof. A very unusual player could in principle trip one of the checks.
- "No sign of a bot" does not clear anyone. A bot that adds random delays, selects groups and uses hotkeys would pass all three checks.
- It says nothing about who is behind an account. A flagged account may be a bot someone else is running, a borrowed account, or something else entirely.
Privacy
The page loads a replay reader into your browser, and that reader does all the work. Replays are not uploaded or stored anywhere.
Credits
- Replays are read with screp by András Belicza, used under the Apache 2.0 licence.
- Pluto is by tscmoo. Its release is at github.com/tscmoo/pluto.
- The analysis and the tool were built with Claude Code, Anthropic's AI coding assistant, running Claude Opus 5.5. A person decided what to look into and checked the results.