[octopols]

• Adds @rivet-dev/claude-code, which runs Claude Code agents as Rivet Actors under the user's own Claude Code login, so no API key is needed
• claudeCodeAgent keeps one persistent Claude Code session per actor, stores every turn in actor state and broadcasts it as a turn event
• runClaudeCode runs Claude Code once and validates structured output against a Zod schema
• Adds a claude-code-research example: a coordinator plans sub-questions, researcher agents investigate them in parallel, and the coordinator writes a sourced report
• Tests run against a fake claude executable, so they need no Claude account

[jog1t]

I think at this point, would be just easier to use Claude's SDK which is preffered way to work with Claude as part of the subscription. btw Claude now adds api credits to any paid account

[octopols]

agreed on the sdk, will switch to it. that should be fine policy-wise, since each user signs in to claude code with their own subscription and uses it for themselves, and we never handle their login or tokens:
https://code.claude.com/docs/en/legal-and-compliance#authentication-and-credential-use

not sure about the credits though, from what I can find only max gets api credits, pro doesn't. pro users are who I built this for, so it'll keep using their own claude code login.

[octopols]

@jog1t switched it to the agent sdk.

• runs through @anthropic-ai/claude-agent-sdk instead of spawning the cli, same public api
• agents only get the tools they list, everything else is denied, and they don't load the user's settings or CLAUDE.md
• executable is now claudeCodePath, defaults to the claude code the sdk bundles
• tests run the real sdk against a fake claude code process, no claude account needed
• cleaned up the example and replaced type casts with typed state

still runs on the user's own claude code login, checked it on a pro plan with no api key.