Choir

Command Line & Claude Code Skill

Everything the GUI does is also reachable from a choir binary on your $PATH. The CLI shares the same database as the app, so keys you stored in Settings work immediately and saved comparison runs show up in the GUI's history. It's small, fast, and designed to slot into shell scripts, test pipelines, and Claude Code as a skill that lets other LLM tools route sub-tasks to a specific model.

One-shot calls

# single model — plain text on stdout
choir ask "Write a regex for IPv6 with zone IDs" --model gpt-5

# multi-model fan-out — JSON array on stdout
choir ask "Suggest 3 commit messages" \
     --models gpt-5,claude-opus-4-7,gemini-3-pro-preview \
     --stdin < diff.txt

# test-pipeline assertion
choir ask "$(cat fixtures/sql.txt)" --model gpt-5 --json --stdin \
  | jq -e '.[0].text | test("SELECT.*FROM users")'

Saved comparison runs (the redo path)

Add --save to a fan-out and Choir persists the run. From there you can layer a judge model on top, retry just the participants that errored, or knit in a response you gathered outside the CLI — all without re-running anything that already succeeded.

# run + save + judge in one shot
choir ask "..." --models gpt-5,claude-opus-4-7,gemini-3-pro-preview \
     --compare-with claude-opus-4-7 --judge "focus on factual accuracy"

# retry only the failed participants (uses each model's CURRENT config)
choir runs add <run-id> --retry-failed --replace

# attach a hand-gathered response as a participant row
choir runs ingest <run-id> --as-model claude-opus-4-7 \
     --text-file response.md --replace

# try a different judge over the same participants
choir runs compare <run-id> --with gpt-5

Hook it up to Claude Code

Drop a single skill file into ~/.claude/skills/choir/ and any new Claude Code session can automatically reach for choir when the user asks for a second opinion, a parallel batch, a capability test, or a specific non-Anthropic model.

1. Build & install the CLI: bash cli/install.sh (defaults to ~/.local/bin; override with PREFIX=/usr/local/bin).
2. Copy or symlink the reference skill: ln -s "$(pwd)/docs/SKILL.md" ~/.claude/skills/choir/SKILL.md.
3. Start a fresh Claude Code session. Existing sessions can use the choir binary right away, but only new sessions auto-load the skill description for triggering.

What the skill does for you: when someone running Claude Code says "have GPT-5 take a crack at this SQL" or "show me how Gemini and Claude differ on this design", Claude shells out to choir with the right flags, parses the JSON, and feeds the responses forward into the rest of the conversation. Comparison summarization stays opt-in — the default is to hand the structured outputs straight back so they can flow into the next step.