Connect AI × DOOM

But can it run Connect AI? Ask Claude a business question, and Claude answers it through the CData Connect AI MCP server. A Doom level is built from what it finds, and every monster in it is a live record from your apps.

• Each room is a data source: Salesforce, Zendesk, Stripe, whatever answers the question.
• Each monster is a record from that source. Aim at one to see it. The biggest monster in each room is urgent.
• Each room looks like its app: a wood-paneled Salesforce office, a Zendesk ops center, a Stripe treasury, a Snowflake data lake, and so on for every app on the account, each with its own monsters.
• Keycards are your permissions: blue for read-only apps, yellow for read + write.
• The red door is data Connect AI blocked, like a lookup by Social Security number. There is no key.

Play it yourself, or let Claude play while a side panel shows every step it takes through Connect AI.

Install

You need a Mac (Apple Silicon or Intel) and about 10 minutes.

1. Homebrew, if you don't have it. Its installer also sets up Xcode's command line tools (git, a C compiler).

2. Claude Code. Install it, then run claude once and sign in.

3. A Connect AI account with a few connections (a CRM, a support desk, billing) and a personal access token: in Connect AI, open Settings (the gear icon) → Personal Access Tokens → Create PAT, and copy the token. It's shown only once.

4. Clone and build. setup.sh installs SDL 1.2 with Homebrew, builds the game (psdoom-ng) and a map builder from source, and downloads Freedoom for the game's art. It takes about five minutes.

   git clone https://github.com/CDataSoftware/connect-ai-doom.git
   cd connect-ai-doom
   ./setup.sh
   ./raid.sh offline        # check the build: sample data, no account needed

5. Log in, once. Your email and token go into the macOS Keychain.

   ./raid.sh login

6. Check the Mac. doctor checks the build, the token, that Connect AI answers over this network, how long each of the level's queries takes, that the game build is current, and the panel's port, and prints what to run for anything missing. Without a token it checks the Mac for offline play.

   ./raid.sh doctor

Play

./raid.sh new                       # Claude briefs on your data (about a minute), then you play
./raid.sh new -q "Which vendors are we late paying?"
./raid.sh replay                    # play the last level again, no new Claude run
./raid.sh replay --claude-plays     # watch Claude play it, again and again (--no-loop: once)
./raid.sh booth                     # replay game after game, for a booth
./raid.sh booth offline             # the same with sample data, no account needed
./raid.sh offline --offline-plays   # Claude plays offline: no network at all, no AI tokens
./raid.sh booth --god-mode --saved-data   # nobody dies, and nothing asks Connect AI

Mode	Claude briefing	Game
new	runs one (about $0.30 of Claude usage)	yes
replay	reuses the last	yes
booth	reuses the last (or none, offline)	one after another
offline	none, sample data	yes
brief-only	runs one	no

The raid takes over the whole screen, for booths and screen recordings: the game on the left two thirds, with Claude's question and the controls under it, and the panel on the right third. When Claude plays, the panel streams each decision and tool call as Claude makes it. Cmd-Tab to another app and the full screen steps aside until you come back to the game; Cmd-Q on the game or the panel ends it at once; LAYOUT=off plays in a normal window instead. The game plays on the easiest skill: you take half damage and pick-ups give twice the ammo. A record only wakes when you come within 384 units in sight or shoot it, so you clear a room a few at a time. Controls: WASD to move, the mouse to turn, click or Space to fire, walk into a door to open it (or press E), Shift to run, P to pause.

Ask a new question mid-game in the box at the bottom of the panel: press P to free the mouse, click the box, type the question and press Enter (or Use the default question to go back to the stock one). The game ends, Claude briefs the new question on the stage while you watch, about a minute, and the next game is its level, with the Options menu's switches as you left them. If Claude can't build a level for it, the last level comes back and the panel says so. It needs a Connect AI account and Claude Code, so an offline game has the box switched off.

Switch things mid-game in the Options menu (Esc → Options):

• Claude plays hands the marine to Claude or takes it back.
• Auto replay (on by default) restarts the level with every record back once it's clear, a few seconds after you die, or when Claude's run ends, then lets Claude play again if it was playing. With it on, a minute with nobody at the keyboard or mouse hands the game to Claude (IDLE_SECONDS).
• God mode: players can't die and never run out of ammo, so nobody's turn ends early.
• Saved data plays the level's last saved records and calls Connect AI for nothing: no refresh, no live read on a kill, no live check at the red door. For when the Wi-Fi is up but slow.
• Offline Claude: Claude's turns use the scripted offline player (below) instead of Claude, so they cost no AI tokens. Flipping it mid-run swaps players where the marine stands.
• Music plays Freedoom's soundtrack, and Sound effects the gunfire, doors and monsters. Turning either off keeps its level on the Sound Volume sliders for when it's back on.

Each is off unless its flag turned it on: --claude-plays, --loop (Auto replay, on anyway unless --no-loop), --god-mode, --saved-data, --offline-plays; Music and Sound effects are on unless --no-music or --no-sound-effects turned them off. In a booth, the switches carry over from one game to the next.

Claude plays offline. When Claude can't run (no network, or claude isn't signed in), on the offline level, or with --offline-plays, a scripted player takes Claude's place. It uses the same Doom tools in the same order (keycards, urgent records first, room by room, then the red door) and writes the same kind of feed to the panel, narrating each kill with the line Claude wrote for that record in an earlier real run of the level, or a line built from the record. Real runs are kept in state/recordings/ for this, so let Claude play each level once while you have a network. The panel labels these runs CLAUDE PLAYS · OFFLINE REPLAY.

At a booth, run ./raid.sh doctor and then new once before the doors open, then ./raid.sh booth. For a family-friendly floor, NOGORE=1 ./raid.sh booth swaps blood for bullet puffs and leaves no corpses: a killed record vanishes in green teleport fog. Each player's game ends with Esc → Quit Game → Y, and the next one starts 10 seconds later with all the records back, on the same full screen (BOOTH_PAUSE changes the wait). Ctrl-C in the Terminal stops it. The screen stays awake while raid.sh runs. On a Mac with no Connect AI account, or no level yet, booth plays offline with sample data; ./raid.sh booth offline does that on purpose.

• On a display or projector, move the Terminal window onto that screen before you run raid.sh: the game and the panel fill whichever screen the Terminal is on.
• If the Wi-Fi is slow or down, each Connect AI call gets 30 seconds (CONNECT_AI_TIMEOUT), and the three rooms are read at once. After that the game plays the records this level last fetched; if it never fetched any, it plays offline and keeps your level for the next game. Kills still count; the panel shows their live lookups as failed until the Wi-Fi is back.
• offline keeps your level. After an offline game, replay plays your real level again.
• Between visitors, ./raid.sh booth --claude-plays keeps Claude playing on its own, all day, and starts the game again if it quits or crashes. A visitor can turn Claude plays off in the Options menu and play; a minute after they walk away, Claude takes over again. If the network drops, the offline player carries on. Each Claude run is up to 60 turns, at about $0.13 per 20. A crash can leave a macOS "quit unexpectedly" alert on screen to click away. Use one screen: with a second one attached, clicking an app there hides the game's full-screen stage until someone clicks the game again.

How it works

Doom never talks to Connect AI directly. Two programs do: Claude Code, before the game, and a small bridge, during it. Both use the Connect AI MCP server with its read-only tools only, so nothing in your apps is ever changed.

 Connect AI MCP server  (https://mcp.cloud.cdata.com/mcp, read-only tools)
       ▲                                  ▲
       │ 1. Claude explores and           │ 2. the bridge re-runs queries
       │    writes the queries            │    while you play
       │                                  │
 brief.py → claude -p              connect_doom.py ◄──── psdoom (Doom)
       │                                  │         list / kill / hit / red door
       ▼                                  ▼
 state/mission.json ──► level      state/*.json ──► side panel (localhost:8666)

1. The briefing builds the level. brief.py runs Claude Code headless (claude -p) with one MCP server, Connect AI, limited to its read-only tools: get_catalogs, get_schemas, get_tables, get_columns, get_instructions and queryData. Claude finds the three apps that answer the question and writes one SQL query for each, returning the same four columns: Id, Label, Detail, Hot. That contract is why any connector can become a room. The rest comes from Connect AI, not from Claude:

• Rooms and monsters come from the rows each query returns.
• Door colors come from Connect AI's permissions for your account on each connection.
• The vault comes from requests Connect AI's Data Security refused during the run.

2. The bridge keeps it live. psdoom-ng is a Doom port that turns processes into monsters. Here it asks connect_doom.py instead of ps, so monsters are records. The bridge talks to the MCP server with a small JSON-RPC client (connect_ai_mcp.py) and calls queryData three ways:

• Every 15 seconds, it re-runs each room's query, so the monsters always match your data.
• On a kill, it reads that one record again and shows the result on the panel.
• At the red door, it re-sends the blocked request, and Connect AI refuses it again, live.

Doom's game loop can't wait on the network, so these calls run in the background and the game only reads local files.

3. Claude can play, too. With --claude-plays, play.py gives Claude a second MCP server next to Connect AI: Doom itself, with the tools look, go_to and attack. Claude reads the records it's about to fight and narrates why each one matters.

More detail: docs/internals.md.

Troubleshooting

• "The game isn't built yet": run ./setup.sh. Its log is in vendor/build.log.
• "No Connect AI token": run ./raid.sh login.
• Empty rooms: your connections returned no rows for that question. Try ./raid.sh new -q "..." with a question about data you have.
• The game is in a normal window and the panel in a browser tab: the game was built before the full-screen layout. Run ./setup.sh again to rebuild it.
• "Rebuilding the game failed": raid.sh rebuilds the game itself after a pull changes it; if that fails, run ./setup.sh and check vendor/build.log.

Contributing

Issues and pull requests are welcome. See CONTRIBUTING.md to get set up; the tests run offline, with no account or game build needed.

License

MIT, see LICENSE. The game itself is psdoom-ng (GPL-2.0), downloaded and built by setup.sh; patches/psdoom-ng-connect-doom.patch modifies it and is GPL-2.0 too. Game art is from Freedoom (BSD). Connect AI and CData logos are trademarks of CData Software, Inc.

This is a fan project. It is not affiliated with, endorsed by, or sponsored by id Software, ZeniMax Media or Microsoft. DOOM is a trademark of id Software LLC.