Security fixes target the latest released minor version. Pre-release candidates receive fixes before release; update to the latest patch within your minor.

AGI IS HERE runs in the browser. Your provider key is stored in localStorage for the app's origin and sent to your chosen provider, OpenAI or Anthropic. Production calls the provider directly; the development server proxies requests locally. Creating and remixing send relevant game source, messages, rendered previews and authoring conversation to that provider. Provider calls use your account and are subject to the provider's data handling policies.

Image generation sends the prompt and any selected reference images to the OpenAI images API using your OpenAI key. Generated images stay in the project and can become native PICTURE or VIEW resources. Details selects the model, quality and size; Generate submits the request. Agent requests and images share a task budget, initially $5. The app displays actual provider-reported usage as spending, for example $0.14 of $5 spent, alongside Stop. An in-flight request finishes before the task pauses after crossing the budget; spending can exceed it. Continue adds another task budget, and Stop ends the task while keeping completed work. For unverified model prices, consult your provider's usage page.

Projects are saved in IndexedDB with a localStorage index. Open Download… from Settings → This game or Home's Game actions. Project file is a private backup containing available chats, notes, images, source, history and saved progress alongside the game. Playable game contains playable resources and public metadata for sharing. API credentials are excluded from both. Downloads are local files; sharing them is a separate action.

• Scripts on the app's origin can access its localStorage, including API keys. Production uses the separate agi.monotio.com origin, so other monotio.com sites cannot directly read its localStorage or IndexedDB. Scripts and dependencies served by the app itself remain trusted. Use a revocable key with an appropriate spending limit.
• Game files and model responses are untrusted inputs. Resource readers, tool schemas and the assembler validate them before use. The worker interprets AGI bytecode through defined opcodes rather than executing it as JavaScript.
• Diagnostic exports can include game content and authoring conversations. Review them before attaching them to an issue.

The shipped CSP and security headers restrict scripts and workers to the app origin and provider connections to OpenAI and Anthropic. Self-hosters should apply these headers; see hosting.

AGI_DEV_KEYS is a development-only convenience for provider keys from the local environment. Vite serves it only to loopback requests in development; production builds exclude it. Enable it only for your own local development session, as described in development setup.

Report through GitHub private vulnerability reporting (Security → Report a vulnerability). Include the affected version, steps, impact and a small reproduction. Keep API keys and private game transcripts out of reports. Security concerns belong in this private channel.

This is a personal project maintained on a best-effort basis, without a guaranteed response time. Reports are read, and fixes and disclosure are coordinated with the reporter in the private report.

Production releases require a PR merged by @joakimriedel into protected main and successful CI. External PRs do not receive deployment credentials. Actions use read-only tokens by default and immutable action references; secret scanning, push protection, Dependabot security updates and private vulnerability reporting are enabled in the repository settings. Account and organization administrators can change these controls; this is not a substitute for securing those accounts.