Inari — AI skill for Claude Code

Inari

**A local-first network of AI agents for the machines you already own.** Start an agent on one computer, move it to another one mid-task, hand a copy to a friend — and keep paying only for what you choose: local models for free, free tiers where they exist, a frontier subscription when the task deserves it.

Inari is built around one idea: **an agent's state is a portable file, not a row in someone's database.** Everything else follows from it — switching model or provider in the middle of a thread, moving a running agent between machines, a phone that supervises without computing, machines that keep working when the network does not.

Where it stands

• Home-made, end to end. The capsule format, the teleport protocol, the mesh, the core (one static Rust binary of about 5 MB, about 6 MB of memory) and the web interface were written for this project. The Python package in src/inari_core is the reference implementation; its tests are the specification the Rust core is checked against.
• Every agent harness through one protocol. Claude Code (through claude-agent-acp), Gemini CLI and OpenFox all speak the Agent Client Protocol, so a thread can change harness or model without losing its history.
• Each machine serves its own interface, the phone reaches any of them, and the Android app shares photos and documents with the model.

How Inari differs from other agent cores

Agent platforms already run coding agents remotely, move them between a laptop and a cloud, or let one app drive several models. Inari does not reuse their code, and it makes different choices:

	Inari	Common approach
Where the agent's state lives	A capsule: a plain tar of JSON and Markdown with a hash-chained journal (ssc/1, docs/ssc.md), readable without Inari	A server-side database of the provider or the app
Moving a running agent	Peer to peer between your own machines: two-phase commit with a fencing epoch, r