FAQ

Is Agent Harness hosted?

The primary workflow is local-first. You choose the model provider and where the work runs; the docs do not require a hosted control plane.

Which models work?

Use a configured provider adapter such as Ollama, LM Studio, Anthropic, DeepSeek, OpenAI, or another OpenAI-compatible endpoint. Run npm run harness -- doctor to catch endpoint and model configuration problems before a task starts.

Does the harness change my repository automatically?

Task execution can change the target repository. Use a disposable checkout or worktree for the first run, keep the task scope narrow, and inspect the diff and verifier result before accepting the work.

What happens when a worker fails?

The result remains inspectable through status and report surfaces. Depending on your configuration, the operator can retry, escalate, or require approval.

Where do credentials go?

Keep credentials in environment variables or the provider's credential store. Never put secrets in task JSON, committed config, screenshots, or support logs.

Is there a public cloud API?

There is no hosted cloud API in the current product. The controller and UI HTTP surfaces bind to loopback by default and can be explicitly exposed for a controlled distributed deployment. For agent clients, use the stdio MCP server and discover its tools at connection time.

What is stable?

The CLI, config, and artifact surfaces are soft-stable. Additive changes are preferred; check the compatibility page before upgrading automation.