oharness

A model-agnostic agent harness for the terminal, built the way Claude Code is: the agent loop, tool execution, context management and permissions are one layer, and the model is a replaceable backend.

npm install -g oharness
export ANTHROPIC_API_KEY=sk-ant-...   # or OPENAI_API_KEY / GEMINI_API_KEY
oharness

Three commands are installed:

| | | |---|---| | oharness | the terminal UI — streaming output, permission prompts, slash commands | | oharness-headless | no UI, no prompts, JSONL on stdout and nothing else | | oharness-web | a browser frontend on localhost |

Switching model mid-conversation (/model sonnet → /model gemini) converts no history, because the history was never held in a vendor's shape.

With no UI there is nobody to ask. oharness-headless denies any call that would need a prompt and reports why — letting them through silently would make an unattended run more dangerous than a watched one, which is backwards. Grant what it needs with permission rules up front, or pass --permission-mode auto deliberately.

Documentation, architecture and the reasoning behind both: https://github.com/Jet1217/OHarness