Visr Evals

Quickstart

Run Harbor Framework-compatible task evals without infra or model keys.

Run Your First Eval

Eval runner

Choose an eval runner, review the prompt, then copy it into that agent.

Agent prompt

# Run Your First Eval

Use this prompt to run a Harbor Framework-compatible task eval with Visr and inspect the results.

## Selection

- Dataset: Runme LLM Judge
- Eval runner: Claude
- Docker: disabled

## Agent prompt

1. Install Visr:

brew install sourishkrout/visr/visr

If needed, follow the downstream runme-harbor dependency installation instructions.

2. Clone the Runme repo and enter it:

git clone --depth 1 https://github.com/runmedev/runme.git && cd runme

3. Run the trial eval:

visr run examples/harbor/datasets/runme-llm-judge --agent claude-code --model sonnet

4. If the trial run succeeds, view the results:

visr view

5. Explain what just happened in plain language. Then inspect .runme/evals/jobs
   and give the user a short results overview: whether the eval completed, what
   was evaluated, and where they should look next.

## Notes

- Good first run for checking agent output with an LLM judge.
- Run the eval locally. The selected agent uses existing subscription access when available.

Free credits

$10

Start your first hosted eval runs on us.

Claim $10 now

How it works

Run task evals locally with the agents you already have installed.
Claude Code, Codex, and others use your existing subscription when available.
Without one, Visr falls back to hosted credits.
Docker runs use hosted credits for strict isolation.

Agent workflow evals

Turn coding-agent runs into a team benchmark.

Visr helps teams test the agentic workflows they actually rely on: Harbor Framework task evals, Claude Code sessions, Codex runs, Cursor habits, OpenClaw playbooks, MCP toolchains, repo instructions, skills, and the prompts that bind them together.

Benchmark real agent work

Capture repeatable task evals from real repositories, then compare agents, prompts, skills, and models against the same evidence.

Score team-specific outcomes

Use criteria that match your codebase: tool use, source quality, command discipline, artifact shape, and regression behavior.

Run local or isolated trials

Exercise the same workflow locally with existing agent CLIs, or switch to Docker when a sandboxed benchmark run matters.

Promote what holds up

Package proven workflows so Claude, Codex, Cursor, OpenClaw, and MCP-backed agents inherit behavior your team has tested.