Q&A

Can I run Claude Code or Codex inside blueclaw?

Yes. Set agent.harness.name to claude-code, codex, antigravity or acp, and point agentCommandPath at the executable. The agent starts as the requester's POSIX user and reaches blueclaw's tools through an MCP catalog published for that turn. blueclaw refuses to start it unless the POSIX boundary is configured, because the agent brings a shell of its own.

How is this different from running the agent in a container?

A container isolates the agent from the machine. blueclaw isolates the people using the agent from each other inside one workspace: different Linux users, different 0700 homes, different groups. The two compose, and terminal.mode sandbox adds bubblewrap where it is available.

What stops the agent from running a dangerous command?

The requester's own permissions. The command runs as an unprivileged user and the kernel refuses what that user may not do. Calls with effects beyond the machine go to a person at the approval gate before they run.

Does it need Linux?

No. It is deployed on Linux, and macOS runs the same POSIX separation tests. macOS has no bubblewrap, so terminal.mode stays native there.

Do I need a chat platform?

No. The api connector accepts a JSON POST and serves replies over HTTP. Postgres is required: the task store and the ledger live there.

Can a harness call tools in parallel?

The bundled loop uses native tool calling, and several calls in one model response run as a batch in order, stopping at the first failure. A call whose input depends on an earlier result still costs a round trip; letting the model write code that calls tools would remove that, but only with a new privileged channel from the requester's process into the runtime, which the POSIX model avoids.

Is there a binary release or a Docker image?

No. Build from source.

Where do I report a security problem?

Follow SECURITY.md and do not open a public issue.