AFS

The new site is live. Here is how we plan to bring ARC into the community, data spaces and everyday applications, with migrations, new products and builders taking part along the way.

OpenAI's Agents API turns the Codex loop into a hosted API: sessions, compaction, tools, sandboxes, plugins. That layer is real. It is still not a computer with identity, durable state, and an owner who can walk away with it. ARC attaches at the seam they already documented.

Not a pivot from blockchain to AI. The harder claim is simpler: roads and utilities can be built early, but a digital city only comes alive when residents who can work around the clock walk in.

Most people will not manage API keys. They will hire agents. The real question is not which model is smarter, but whose ground the worker stands on when the job is done.

arc skill ships a small catalog of built-in how-to booklets and can drop them into your agent skill folders. That is useful—and incomplete next to gh skill or AFS’s own .skills trees. Here is the plain map, a short how-to, and the open design gap.

Many companies are building compute for agents: sandboxes, VMs, even digital workers. They answer where untrusted code should run. What agents still lack is a machine that can last: identity, state, constrained authority, verifiable history, and the right to leave. We call that machine ARC.

The runtime for Blocklets. It gives a developer a place to run an application described as a Blocklet, together with the resources that Blocklet declares it needs.

The functional, composable, and typescript-first AI Agent framework for real-world LLM Apps.

ARC is the runtime for Blocklets. It gives a developer a place to run an application described as a Blocklet, instead of making the runtime, the application, and its resource connections one unexplained deployment.

AUP (Agentic UI Protocol) lets an application express what is shown and what actions are available. A runtime then renders the parts that fit its device capabilities, instead of treating one screen's pixels as the application itself.

Useful context is more than a prompt. It includes the documents, tools, and permitted actions that are relevant to the task. If those boundaries are vague, a capable model still has to guess what it may use.

Every agent, app and process sees not the global world but a small world projected specifically for it. What is outside it is not "access denied", it does not exist.