This directory holds Berth's agent framework: code that drives an LLM against the sandbox, rather than the sandbox itself. The sandbox (the berth.yml manifest, the kernel policy compiled from it, the apps that run under it, and the evidence that it held) lives in packages/, apps/, spec/ and bench/.
| Package | What it is |
|---|---|
agents |
@berthos/agents: Computer, Agent, Crew, LLM providers, sessions, guardrails, tracing, an MCP client, A2A and evals |
agents-python |
berthos-agents: the Python Agent and Crew, which connect to a running sandbox |
seam-claude-agent-sdk, seam-openai-agents |
Adapters that hand Berth tools to the Claude Agent SDK and the OpenAI Agents SDK |
Everything here is built and tested with the rest of the repo, and it is frozen: bug and security fixes only, no new features (why).
None of it is published to npm or PyPI. Releases ship the sandbox only. Use these packages from a clone of this repo. The CLI commands built on the framework (berth agent run, berth crew run, berth eval) work from a clone too; an installed @berthos/cli tells you so instead of running them.
To put your own agent on top of Berth without this framework, use berth mcp, the toAiSdkTools / toLangChainTools adapters, or the HTTP RPC bridge.