Kodosumi

Kodosumi runs your agents on Ray Serve, the same scheduling layer behind large production workloads. A single app starts on a laptop and later spreads across a full cluster without a rewrite of the agent logic. That matters when jobs run long or arrive in bursts, because Ray places each one on an available worker instead of blocking a single process.

The admin panel is the part most agent frameworks skip. It lists your registered services, shows a timeline of every execution and records errors as they happen. An analytics view adds throughput and failure counts, and a separate dashboard tracks Masumi revenue. You get the operational picture without building a custom logging stack first.

Kodosumi doesn't ask you to adopt its own way of building agents. It wraps whatever you already use, whether that's a plain Python function, a CrewAI crew or a call to a self-hosted model. Because it stays out of the agent logic, you keep your existing tools and can swap models or libraries later without touching the framework.

You describe the input an agent needs with a small set of form fields, and the framework renders the HTML and validates input on the client side. Agents can send and receive files too, using chunked transfer for uploads and downloads. That turns a bare function into a page a non-developer can actually use.

Some jobs need a person in the middle. Kodosumi supports a lock and lease pattern that pauses an agent, hands control to a human for a decision or an edit, and then resumes the run. It's the missing piece for review steps, approvals and agents that act on sensitive data.

For agents sold as a service, Kodosumi connects to the Masumi payment system on Cardano and speaks the MIP-003 protocol for discovery and job management. You get a payment dashboard with revenue and failure tracking, so an agent can charge per job without you building a billing layer by hand.