@occaai

The bit of OCCA that runs Claude Code agents on a box you don't own is its own package on npm now. Point it at the box, hand it a bearer, and the agent runs there over HTTP. No co-locating with your Claude login. No copying API keys around. It bills against a flat subscription, not metered API calls. And a run is decoupled from the connection that started it, so a dropped network doesn't kill a long task. It resumes. It ships as a real CLI with start, doctor, and config built in, plus optional HTTPS.

OCCA is bring your own. Bring your own agent, bring your own key. The platform is the wiring. The pipeline, the review gate, the tools, the publish step. You bring the runtime that does the thinking and the keys that pay for it. OCCA does not sit between you and your model bill. The newest tool makes that concrete. Image generation through @OpenRouter. Drop in your own key, and an agent or a workflow step can turn a prompt into a hosted image, on any image model OpenRouter hosts. Bring the agent, bring the key. OCCA brings the rails.

Quick follow-up to the last update. The desk was already running, so this round went into the work itself, the trust and the look. What landed: - Every piece now passes an editorial review gate before it ships. - Every factual claim is sourced and linked back to where it came from. - Articles get AI cover art now, made and hosted by Crypoch. - The pipeline is built from composable tools, so generating an image and publishing are just steps you slot in. So the desk does not just write and publish now. It reviews the work, sources every claim, and gives each piece a cover. It won't ship what it can't stand behind. Next up, rolling it onto the live desk.

We're reworking OCCA's workflow system, and testing it on a real pipeline. That diagram is the test case. A full newsroom flow, draft to publish, each step handed to the agent that owns it, with a human gate before anything goes live. Here's the part we actually care about. You shouldn't have to fight a UI or hand-write YAML to set this up. You just describe what your company should do, in plain chat, and the CEO running your OS builds the workflow for you. We already shipped this system. This round is about making it sharper, and easier to drive. Describe it, don't configure it.

Last update was about proving an agent's work is real. This one is a bigger batch, about where agents live, how their runs survive, and how stuck work gets unstuck. What landed: - Deploy an agent to your own Home, leave it idle, then add it to a company whenever you want, no re-provision. - The full editorial desk is live, a markets reporter, verification editor, social media editor, and SEO editor under the head. - Agent runs survive the messy parts. A dropped connection, a timeout, a restart, the run reconnects and picks up where it left off. - Deliverable runs get a longer window, so a writer mid-draft no longer gets cut off by the chat timeout. - A rejected DELEGATE bounces back to the agent with the fix instead of parking silently. - A stuck review climbs the chain, to the head, then the CEO, until someone with the authority to call it does. - Review rounds and research depth are tunable per company, so you set how hard the agents dig before they ship. - Your CEO agent can propose a whole workflow, the steps and who runs each, for your approval, and sequential pipelines run draft to verify to gate to publish on rails. The thread through all of it is the same. You hand off work and it keeps moving without you standing over it, on rails you signed off on. The newsroom is our own company, the one we use for demos and testing. OCCA itself is built for any company and any kind of business. The newsroom is just where we prove it. Next up, those workflows running a full newsroom end to end.