Harness & LLM eval

Benchmark harnesses and models on your real work.

Launch identical agents across Claude Code, Codex, Hermes, and OpenClaw — same prompt, same capabilities, different brains — and compare them on the tasks you actually care about.

A matrix in minutes

Spin up an agent per harness × model combination in one click each. No infra, no API wiring, no eval framework to stand up first.

Identical conditions

Every contender gets the same system prompt, capabilities, and channels, so the comparison is about the harness and model — nothing else.

Swap without resetting

Change an agent's model mid-run. History stays intact, so you can A/B how different models continue the same long task.

Pay only while they run

Park the losers for free the moment you've seen enough. Keep the winner working.

Drive it from the CLI

Script the whole matrix — create, list, swap, park — straight from the terminal.

Real tasks, real channels

Evaluate on your actual workflows over web chat, Slack, or Telegram instead of synthetic benchmarks.

One click to launch. Zero lock-in after.

Launch an agent