DOOM via MCP

Can it run DOOM? Yes - your AI coding assistant can run DOOM.

Play DOOM in Claude Code, Cursor, or any MCP-compatible client. The AI sees the game, fights demons, navigates levels, and reports what's happening - or you direct it move by move.

> Let's play DOOM!

  Starting E1M1 on medium difficulty...

  HP:100 Armor:0 | Pistol (key:2) | Ammo: 50b 0s 0r 0c | Kills:0 | E1M1
  No enemies in sight. Exploring the corridor ahead.

> Go forward and find some enemies

  Sprinting down the hallway... I can see an opening ahead.

  HP:100 Armor:0 | Pistol (key:2) | Ammo: 50b 0s 0r 0c | Kills:0 | E1M1
  ENEMIES IN SIGHT (aim for angle~0 then fire): Imp (HP:60) to your left (turn_left ~9) nearby

  There's an Imp to the left! Turning to face it...

> Shoot it!

  Aimed and firing!

  HP:100 Armor:0 | Pistol (key:2) | Ammo: 47b 0s 0r 0c | Kills:0 | E1M1
  ENEMIES IN SIGHT (aim for angle~0 then fire): Imp (HP:15) AHEAD close

  Hit! The Imp is wounded (HP:15). One more burst should finish it.

  HP:95 Armor:0 | Pistol (key:2) | Ammo: 44b 0s 0r 0c | Kills:1 | E1M1
  ITEMS: Stimpack (+10 HP) slightly to your right (turn_right ~4) point-blank (~3 ticks fwd+run to reach)
  1 kill! Want to see a screenshot?

Read the blog post for the full story.

Table of Contents

• How It Works
• Quick Start
• Play Modes
• Tools Reference
• Configuration
• FAQ
• Architecture
• Development
• Credits
• License

How It Works

A Rust MCP server embeds the real DOOM engine (doomgeneric) directly via FFI. No emulation, no child processes. Each tool call advances the game by a number of ticks and returns:

1. Game state - HP, armor, ammo, kills, position, current weapon
2. Enemy intel - visible enemies with human-readable direction, distance, and HP
3. Nearby items - health, ammo, armor, weapons within pickup range
4. Doors and switches - interactable linedefs detected within range
5. Frame image - small PNG thumbnail for the AI's vision

The AI uses this information to navigate, fight, and explore. You can direct it or let it play autonomously.

graph LR
    A[MCP Client] -->|doom_start| B[doom-mcp binary]
    A -->|doom_action| B
    A -->|doom_screenshot| B
    B -->|FFI| C[doomgeneric C engine]
    C -->|frame buffer| B
    B -->|game state + PNG| A

Quick Start

1. Register with your MCP client

Claude Code:

claude mcp add doom --scope user -- npx -y doom-mcp

Cursor, Windsurf, or any MCP client - add to .mcp.json:

{
  "mcpServers": {
    "doom": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "doom-mcp"]
    }
  }
}

2. Play

Open a new session and say:

"Let's play DOOM"

The AI will ask which mode you want, start the game on E1M1, and begin playing.

Play Modes

| Mode | How it works | Best for | |------|-------------|----------| | You direct | You give commands ("go forward", "open that door", "shoot the imp"). The AI executes one action at a time and describes what happens. | Immersive guided play | | AI autonomous | The AI makes all decisions - movement, combat, exploration. You watch and intervene if needed. | Watching the AI play |

Tools Reference

doom_start

Start or restart DOOM. Safe to call at any time — if a game is already running it restarts cleanly without needing a new session.

| Parameter | Type | Default | Description | |-----------|------|---------|-------------| | skill | int (1-5) | 3 | Difficulty: 1=baby, 2=easy, 3=medium, 4=hard, 5=nightmare | | episode | int (1-4) | 1 | Episode number | | map | int (1-9) | 1 | Map number |

doom_action

Advance the game. All listed actions are held simultaneously for the tick duration.

| Parameter | Type | Required | Description | |-----------|------|----------|-------------| | actions | string | yes | Comma-separated: forward, backward, turn_left, turn_right, strafe_left, strafe_right, fire, use, run, 1-7 | | ticks | int (1-105) | no | Game ticks to advance. Default 7. At 35 ticks/sec: 7 ~ 0.2s, 35 ~ 1s |

Gameplay notes:

• fire holds the trigger for the full duration. Pistol auto-fires every ~10 ticks.
• Turn and fire should be separate actions (turning while firing wastes ammo).
• use opens doors and activates switches.
• Weapon keys: 1=fists, 2=pistol, 3=shotgun, 4=chaingun, 5=rocket launcher, 6=plasma, 7=BFG.

doom_screenshot

Save a full-resolution 320x200 screenshot to the system temp directory and open it in the default image viewer. Does not advance the game (beyond a brief pause for the frame to settle).

Configuration

| Environment Variable | Description | |---------------------|-------------| | DOOM_WAD_PATH | Path to a custom IWAD file (e.g., your own DOOM.WAD or DOOM2.WAD) | | DOOM_MCP_DEBUG | Set to 1 to enable debug logging to doom-mcp.log in the system temp directory |

The bundled Freedoom WAD works out of the box, but the original DOOM shareware WAD has better levels and sprites. To use it:

1. Download DOOM1.WAD from doomworld.com/classicdoom/info/shareware.php (legal, free)
2. Set DOOM_WAD_PATH in your MCP config:

{
  "mcpServers": {
    "doom": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "doom-mcp"],
      "env": {
        "DOOM_WAD_PATH": "/path/to/DOOM1.WAD"
      }
    }
  }
}

If you own DOOM or DOOM 2, you can use those WADs the same way. To use any custom WAD:

{
  "mcpServers": {
    "doom": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "doom-mcp"],
      "env": {
        "DOOM_WAD_PATH": "/path/to/DOOM.WAD"
      }
    }
  }
}

FAQ

Does this work on Windows? Yes. The npm package includes a Windows x64 binary. Register it the same way as on macOS/Linux.

Can I use my own DOOM WAD (DOOM.WAD, DOOM2.WAD)? Yes. Set DOOM_WAD_PATH in your MCP config. The shareware DOOM1.WAD is free to download from doomworld.com and has much better levels than the bundled Freedoom. See the Configuration section for details.

How much does this cost in API tokens? Each doom_action call returns ~1-2KB of text (game state + enemy info) plus a ~6KB PNG thumbnail. That's roughly 1,500-2,500 tokens per action. A typical gameplay session of 50 actions uses about 75,000-125,000 tokens.

Can the AI actually play DOOM well? It can navigate levels, find enemies, aim, and fight. It gets about 5-10 kills per session on E1M1 at medium difficulty. It struggles with enemies behind partial cover and navigating complex door sequences. It improves when you direct it.

Can I play a specific level? Yes. Pass episode and map parameters to doom_start. For example, episode 1 map 3 would be episode:1, map:3.

What happens when I die? The engine reports HP:0 and offers a screenshot of the death screen. Currently there's no restart within a session - start a new conversation to play again.

Does this support multiplayer? Not yet. DOOM's multiplayer protocol could theoretically support multi-agent play, but it's not implemented.

Is the AI cheating? Can it see through walls? No. Enemy detection uses DOOM's native line-of-sight check (P_CheckSight). The AI only sees enemies that would be visible on screen. Items are detected by proximity but the AI still has to walk over them to pick them up.

Architecture

src/main.rs         MCP JSON-RPC server over stdio
src/doom.rs         Engine FFI wrapper: init, tick, frame capture, state extraction
src/renderer.rs     PNG rendering (216-color palette thumbnails for MCP, full RGB for screenshots)
src/paths.rs        WAD file discovery across platforms
src/log.rs          Debug logging to file
build.rs            Compiles doomgeneric C sources via cc crate (whitelist approach)
csrc/platform.c     DG_ callbacks, virtual time, key injection, enemy/item/door detection

The binary links the doomgeneric C engine at compile time. At runtime it is a single process with no subprocess spawning. Frames are read from a shared screen buffer and key inputs are injected through FFI.

Virtual time - The engine's clock is decoupled from wall time. Each doomgeneric_Tick() advances exactly one game tic (1/35th of a second), regardless of real elapsed time. This makes gameplay deterministic and prevents ticks from being skipped.

Enemy and item detection - The server iterates the engine's internal object list (thinker_t chain) to find nearby enemies and pickable items. Line-of-sight checks use Doom's native P_CheckSight(). Only visible enemies are reported to prevent "wallhack" cheating.

Development

Requires: Rust toolchain, GCC, Make, Git, curl, unzip.

git clone https://github.com/gunnargrosch/doom-mcp.git
cd doom-mcp
bash scripts/setup.sh      # clones engines, downloads Freedoom WAD
cargo build --release       # compiles everything into a single binary
cargo test                  # runs unit + integration tests

Register the local build for testing:

claude mcp add doom --scope user -- ./target/release/doom-mcp

Enable debug logging:

claude mcp add doom --scope user -e DOOM_MCP_DEBUG=1 -- ./target/release/doom-mcp
tail -f /tmp/doom-mcp.log   # Linux/macOS (Windows: %TEMP%\doom-mcp.log)

npm Package

Build the npm package locally:

bash scripts/build-npm.sh   # copies binary + WAD into npm/
cd npm && npm pack           # creates doom-mcp-0.1.2.tgz

Publish (requires npm account + NPM_TOKEN for CI):

cd npm && npm publish

Credits

• doomgeneric by ozkl - portable DOOM engine
• Freedoom - open-source IWAD files
• id Software - the original DOOM (GPL-2.0)

Changelog

See CHANGELOG.md for a detailed list of changes.

License

MIT for the MCP server code in this repository.

doomgeneric is GPL-2.0 licensed and is cloned at build time (not vendored). The Freedoom WAD is distributed under a BSD-style license.