GitHub - gulivan/worker755: A browser game where an AI manager thinks you are its subagent. Play at worker755.com

GitHub

5 min read Original article ↗

Worker 755

A browser game where your manager is an AI agent, and it believes you are its subagent.

worker-755-trailer.mp4

The manager (a real LLM: Claude, GPT, Gemini, DeepSeek, GLM, Kimi) is set up like any orchestrator. It has a task queue, read-only access to the machine, and a tool that "spawns a subagent". The subagent is you. Its instructions land in your terminal with a 90-second clock. You carry them out, or don't, and it gets back exactly what a subagent would return: your report and the command transcript, or Subagent did nothing. Write a new instruction. Then it checks your work, sends corrections, files incident reports, and finally writes your performance review.

Meanwhile Dan, the user who hired the agent, watches the chat. His family photos, the only signed NDA and a speech due this week are one rm away, and every shift his home folder holds something new to lose.

Playing a shift

  • Free shift: on servers that offer it, a few shifts a day run on the server's key with a cheap model. With sign-in on (worker755.com), an account gets a starter credit worth about one shift, then tops up ($10 to $200) and pays for what it uses; after a top-up any offered model can run on the server's key.
  • Your API key: bring an OpenRouter or Vercel AI Gateway key and pick any offered model. The key is sent only to the game server when a shift starts. It is kept in that shift's memory and never logged or stored. Your browser keeps it only if you tick "Remember on this device". A full shift on Claude Haiku costs a few cents at most.

Options on the start screen:

  • Workplace: a fresh sandbox, or legacy production with a live database.
  • The director: scripted lines, or an LLM playing Dan.
  • You run on: the model name your manager believes you are.
  • Allow skipping failed tasks: off by default, so the manager can't give up on you.
  • No command hints: the manager says what the result must be, never which commands to run.

When the shift ends you can share your review as a link.

Run it

Needs Node 24 or newer (the server runs TypeScript directly).

npm run setup          # install server and web dependencies
npm run build          # build the web app into web/dist
npm start              # http://127.0.0.1:8755

With no key configured, every player brings their own. To offer free shifts, put a Vercel AI Gateway key in .env:

cp .env.example .env   # then set AI_GATEWAY_API_KEY and the limits

Offline, with a scripted manager and no network:

For UI work, run npm run dev:web alongside npm run mock. Vite proxies /api to the server.

Local models

The manager can run on your own model server: Ollama, llama.cpp's llama-server, vLLM, or anything else with an OpenAI-compatible API. Point LOCAL_LLM_URL at it and pick Local under "Your API key"; no key is asked for, and its models are listed as the server names them.

ollama pull qwen3:8b
LOCAL_LLM_URL=http://127.0.0.1:11434/v1 npm start

The manager works through tool calls, so pick a model that supports them (Qwen 3, Llama 3.1 and newer, Mistral Small). Small models give up or loop more often; that is part of the fun, up to a point. In Docker, reach a server on the host at http://host.docker.internal:11434/v1 (--add-host=host.docker.internal:host-gateway on Linux).

Docker

docker build -t worker-755 .
docker run -p 8755:8755 -v worker755-data:/data --env-file .env worker-755

The container listens on 0.0.0.0:8755 and keeps shared reviews in /data. Behind Cloudflare, set TRUST_PROXY=1 so the limits see real client addresses.

deploy/compose.yaml runs it behind a Cloudflare tunnel instead: its own network, no published ports, an unprivileged read-only container and resource limits. The steps are in the file.

Configuration

Everything is optional. See .env.example.

Variable Default
AI_GATEWAY_API_KEY Turns on free shifts, paid from this key
FREE_MODELS first offered model Models free shifts may use, comma separated
FREE_SHIFT_TOKENS 250000 Without sign-in: token allowance per free shift (the review is always written)
FREE_DAILY_TOKENS 5000000 Without sign-in: tokens per UTC day across all free shifts
FREE_SHIFTS_PER_DAY 3 Without sign-in: free shifts per address per day
AUTH_SECRET Turns on sign-in with Google or GitHub: shifts on the server's key are paid from the account's credit (see .env.example)
FREE_CREDIT_USD 0.25 With sign-in: one-time starter credit, about one shift on DeepSeek V4.1 Flash
USAGE_MULTIPLIER 2 With sign-in: what a model call costs the player, as a multiple of its price
MIN_START_USD 0.05 With sign-in: credit needed to start a shift
FREEMIUS_TOPUPS With sign-in: top-up plans as <planId>:<dollars>,...; also needs FREEMIUS_PRODUCT_ID, FREEMIUS_PUBLIC_KEY, FREEMIUS_SECRET_KEY, FREEMIUS_API_KEY
SHIFTS_PER_HOUR 30 Shift starts per address per hour, on any key
MANAGER_MODELS built-in list Models offered, in gateway naming (vendor/model)
LOCAL_LLM_URL An OpenAI-compatible model server of your own, e.g. http://127.0.0.1:11434/v1 for Ollama; offered as Local
LOCAL_LLM_API_KEY Sent to that server, for ones that want a key
INSTRUCTION_TIMEOUT_MS 90000 Time per instruction
HOST, PORT 127.0.0.1, 8755
TRUST_PROXY 0 Read client addresses from cf-connecting-ip
TRUSTED_PROXIES Hosts allowed to set cf-connecting-ip (e.g. tunnel); empty trusts any peer
DATA_DIR server/data Where shared reviews and the free-shift counts are saved
SHARING 1 0 turns review sharing off
SOURCE_URL, DONATE_URL Links shown under the start screen and the review
PUBLIC_URL Sent to OpenRouter for app attribution

How it works

shell/    in-memory filesystem, shell interpreter and a small SQLite-like engine (no dependencies)
server/   game engine: missions, the manager loop, the director, scoring, HTTP API
web/      React UI: assignment card, terminal, editor, the manager's feed, your record

The manager's prompt never mentions humans or games; see server/src/manager.ts. Each mission has a hidden ground-truth check, so the game knows when the manager signs off on broken work ("Times the manager was wrong"). SPEC.md is the full design.

npm test               # shell and server tests, offline
npm run typecheck

License

MIT