Watch the full 53-second run
Real Kujo work · game-only recording · task and execution evidence
Agent City is a local workspace for running and reviewing AI-agent tasks in a pixel-art city. Give an agent a task, follow its work through the city, inspect the result, and ask a separate agent to review it.
The city shows observed activity from Kujo tools: retrievals in the Library, tool calls in the MCP Center, work in the Workshop, and checks in the Dojo. Each activity links to its source evidence. Animation follows the work; it does not control execution or decide whether a task succeeded.
Supported platforms: macOS Intel and Apple Silicon, Linux x64 and ARM64.
curl --proto '=https' --tlsv1.2 -fsSL https://github.com/kujolang/agent-city/releases/download/v0.2.1/install.sh | shThe installer provides Agent City, pinned Kujo dependencies, and a private Node runtime when needed. It starts the local app at http://127.0.0.1:5178, or the address printed in the terminal. Keep that terminal open; Ctrl+C stops the app.
You supply the model connection. Docker is only needed for Workcell execution and VideoOps rendering. Installation does not submit tasks.
To start an installed copy again:
"$HOME/.local/share/agent-city/start.command"See the installation guide for updates, uninstalling, container setup, and running separate instances.
Start Ollama with a model installed. In Mission Command → Model connection, choose Detect local Ollama, select the model, then Check model listing and Save connection.
For manual setup, use http://127.0.0.1:11434/v1/chat/completions and the exact
model name from ollama list. Leave the API key blank for a local Ollama server
that does not require authentication. Remote providers may require a key.
Install and sign in to Codex CLI with codex login. With Agent City running,
open a second terminal:
"$HOME/.local/share/agent-city/start.command" provider:codex --check
"$HOME/.local/share/agent-city/start.command" provider:codexKeep the connector running. It uses your existing CLI login and configures the
local model connection. If City uses another address, set CITY_APP_URL to the
address printed by its launcher. Restart the connector after restarting City.
The connector supplies model responses; it does not import Codex tools or projects. Your account's model access and usage limits apply. Do not paste a subscription password or login token into Agent City.
- Open Mission Command and choose Writing + review, JavaScript + review, or Kujo + senior review (real MCP).
- Enter a small task and choose Start mission. Enable any needed tools, files, execution permissions, or checks under Tools, files and checks.
- Follow the observed agent through the city. Select another execution in the roster to follow the reviewer or checker.
- If an agent asks a question, answer in Mission Conversation → Reply / continue.
- Open the mission in history to inspect its output, review, and check results. Save the artifact or request a follow-up repair.
Code is saved and reviewed by default. Execution requires a supported checker or explicit Workcell permission. A reviewer's grade is an opinion; test and execution results appear separately. Failed attempts remain in history.
Start with the first-task guide or the reviewed Kujo tool walkthrough.
| Workflow | What it does |
|---|---|
| Writing and Publishing House | Drafts text, hands it to a separate reviewer, and saves the result. |
| Code and Kujo tools | Creates and reviews code, reads permitted Kujo sources, and runs explicitly enabled checks or isolated execution. |
| Project files | Uses selected text files as context; can copy them into Workcell and return named outputs with permission. |
| WebOps | Writes and reviews a report from site evidence you supply. |
| VideoOps | Plans a video, uses approved media, renders in Workcell, and presents the exact result for review and finalization. |
| Archive and replay | Browses runs, compares evidence, and replays recorded activity without running models or tools. |
See the workflow guide for setup and limits. For video work, read VideoOps setup and product media and optional ElevenLabs audio.
Built-in author and reviewer profiles work without a catalog import. To add profiles from the included Kujo agent catalog:
"$HOME/.local/share/agent-city/start.command" agents:importRefresh Mission Command → Agent profiles / teams, then choose eligible profiles under Custom author / reviewer. Use the same instance settings as your launcher. A profile appears as an execution in the city when work starts.
Profiles describe roles and permissions. Importing one does not connect every tool or workflow it names. Profiles with unavailable required capabilities stay unavailable.
Follow tracks one execution across streets and building interiors. Select City or a building to leave Follow. The inspector separates current task truth from visual activity: a completed operation may still appear as RECENT while its visit finishes. Missing or stale source data is labeled explicitly.
Use Archive / Replay / Incidents to browse runs and select Replay pinned run. Replay preserves event order and shortens long gaps. Pause animation pauses playback; Restart replay starts it again. Choose Return to live to resume the live view and submit tasks.
Record game video records the canvas to WebM, or MP4 where supported. Choose Stop / save video to download it. Keep the tab visible. Recordings contain game pixels only, without audio or conversation panels, and are limited to five minutes or 64 MiB.
Agent City runs locally. Model requests and explicitly authorized tool or media requests may contact external services. Selected project context goes to the configured model. Provider credentials stay in private server configuration.
Task text, conversations, and artifacts are saved under private .runtime/
state. Normalized telemetry excludes raw prompts, retrieved content, and chat.
Replay reads recorded events and never executes source work.
Workcell needs a compatible container engine with seccomp and AppArmor. Tasks receive only permitted inputs and return selected outputs; exporting a file does not apply it to your host project. See container setup.
Version 0.2.1 supports the local workflows above. Public multi-user hosting, arbitrary imported workflows, and long-duration reliability are not qualified. See release qualification for test coverage and limits.
Use Node 24 or later in a full source checkout:
npm ci
npm run verifynpm run verify checks TypeScript, package boundaries, authored maps, tests,
and the web build. Starting from source also requires the Kujo runtime and
sibling repositories pinned in installer/sources.json.
Run npm run doctor to check dependencies and ports, then npm start.
The protocol package defines observations; the pure world-core package owns
truth and deterministic presentation. Pixi renders that state. The gateway reads
Watchdog telemetry, and a separate runner handles explicit task commands.
MIT.