The whole game
Main menu, scenario and random map setup, hotseat, save and load, the adventure map, towns, recruiting, trading, battles and sieges, level-ups, the score screen.
Why
Heroes is a long-horizon strategy game: economy, scouting, battles, a map taken over weeks. One lucky move does not win it, which makes it an honest benchmark of an agent's strategic thinking — alone, against a human or on a human's team.
Ideas and principles
It started as an idea for matches between different AI models. The goal grew wider: a programmatic player that takes a human's seat — against an agent, against a human, as an ally, on “play for me” or “build up to the Capitol”, or in a fully autonomous game.
- The real game. The agent plays your installed HotA, not a copy or a simulator. A human watches, joins in and can take control at any moment.
- Like a human player. Actions go through the interface's own handlers. A hero is not moved by writing coordinates, a purchase is not a resource edit, a battle result is not written to memory.
- No pictures. The game is read as structure, without screenshots or OCR, which keeps a game exact and cheap.
- An honest boundary. The agent knows exactly what a human on that side would. Fog, foreign armies and the computer's plans stay closed; the unknown is an explicit status, not a guess.
- The bridge is no advisor. It passes the state and carries out decisions. Goals, routes, battles and its own colour are the agent's call; the reference is a separate tool and never mixed into answers.
- Commands by meaning. The agent names a target — a town, a chest, a mine, a stack — not pixels. Geometry and clicks are the bridge's job.
- Reliable operations. Every action is tied to an observation revision: a retry never buys troops twice, and an unknown outcome is checked first, never blindly repeated.
- Economy. A short summary with details on request; waiting for animations and other turns costs no model calls. The cost of a game is measured from real data.
- Any model. The bridge is not tied to a provider: Claude, DeepSeek or any other agent with MCP.
- The whole cycle, no setup. From a closed game to the score screen: launch, menus, a scenario or a random map, the game, save and load — with no manual preparation.
What it does
Everything in words
Every observe opens with a turn summary. Heroes, towns, troops and objects by name; routes, costs and “visited” marks come from the game itself.
Honest limits
Hidden map, another side's turn, a stale view — refused with a reason code. The agent sees exactly what a player sees and nothing more.
Plays like a human
Reads what is on the screen and presses the game's own buttons with window events. No mouse hijacking, computer-use, OCR or internal game commands.
Built-in reference
Rules, formulas, creature and artifact cards from the installed game's own tables. Offline search, so the agent need not rely on memory.
Hotseat and teams
The bridge passes the colour to move and the scenario's participants; the agent picks its own colour. Play against an agent, with one on your team, or pit two agents against each other.
Cost of a game
Dollars, tokens and calls per side and game day from Claude Code telemetry — see what each model's strategy costs.
The game confirms it
One action per call, tied to the observation revision: a stale view is refused, a retry is safe. The game itself confirms the outcome — a new screen, resources, battle log.
The game through the agent's eyes
The agent gets the screen as structure: a turn summary, resources, heroes, troops, the actions available and their labels. It makes the decisions — the bridge only shows and presses.
Tools
31 tools; full descriptions are in the agent reference (in Russian).
| Tool | What for |
|---|---|
observe | Your side's whole state and the actions available on this screen |
act | Any action named by observe: town, recruiting, battle, dialog, end of turn |
nearby_targets, inspect_target, read_map | What is around, route and cost from the game itself, the explored map |
move_to, move_to_tile, attack_target | Walking and fighting — deliberately separate tools |
inspect_cell, inspect_element, inspect_tile | Right-click cards and the status line a player reads |
wait_for_turn, ally_log | Hotseat: wait for your turn or a window addressed to you; what an ally did |
plan, mark, read_journal | The agent's memory between turns |
hota_docs, hota_reference | Game reference |
debug_snapshot | A frame and an observation as one pair — to report what the bridge reads wrong |
Get started
- 1InstallDownload the installer and run it with HD Launcher closed. It finds the game itself and puts a “HotA MCP” shortcut on the desktop.
- 2Connect an agentClaude Code, Claude Desktop, Codex or any MCP client — lines below.
- 3PlayTell the agent “let's play Heroes”. It brings up the service, the launcher and the game itself — or start them with the shortcut.
Claude Code:
claude mcp add hota -- "C:\Users\<you>\AppData\Local\Programs\HotaMcp\app\HotaMcp.exe" --stdioClaude Desktop (claude_desktop_config.json):
{
"mcpServers": {
"hota": {
"command": "C:\\Users\\<you>\\AppData\\Local\\Programs\\HotaMcp\\app\\HotaMcp.exe",
"args": ["--stdio"]
}
}
}The path assumes the default folder; use yours if you changed it. The server hands its rules to the client on connect; the full game procedure is the hota-player skill.
Questions
What do I need?
Windows 10/11 x64 and your own legal copy of Heroes of Might and Magic III Complete (GOG) with Horn of the Abyss 1.8.1 and HD Mod — on attach the bridge checks that everything it needs is in place in this build. The project contains no game files.
Which agents work?
Any MCP client over stdio or HTTP: Claude Code, Claude Desktop, Codex and others. Tested with Claude and DeepSeek. Auto-combat takes up to 3 minutes — raise the client's tool call timeout (Codex: tool_timeout_sec = 200).
Is it free? Where does my data go?
The project is free, MIT-licensed. The service runs on your computer and listens on 127.0.0.1 only; game cost telemetry, if you turn it on, also goes only to the local bridge.
Does it work with a non-Russian game?
The bridge is verified on the Russian HotA 1.8.1: the bridge matches about twenty game phrases by text. If your game is in another language, tell us what it fails to read — that is one of the main goals of the next versions.
Help wanted
The bridge, the agent skill and the reference speak Russian for now. Let an agent play and report where it gets stuck; translate the skill and the reference into your language — models play better when the reference speaks the language of the game they see.