rubric-chat

A strict 0–100 score for AI conversations.

Auto-discover terminal-agent sessions on disk or upload an export — Rubric grades the prompting craft on six dimensions (specificity, context, structure, iteration, scope, meta-prompting), names an archetype, and gives you a public score card you can share.

Install

npm i -g rubric-chat

Requires Node ≥ 18.

Usage

# Interactive wizard — auto-login, picks a session, scores it, opens the report.
rubric-chat

# Common subcommands
rubric-chat status            # who am I, what plan, where are the URLs pointed
rubric-chat login             # browser-callback login (or --no-browser to paste a token)
rubric-chat logout
rubric-chat list              # show discovered local sessions
rubric-chat rate <id>         # score a specific session id
rubric-chat rate --file path  # score a JSON file (ChatGPT or Claude.ai export)

# Output modes
rubric-chat rate <id> --json   # machine-readable
rubric-chat rate <id> --quiet  # exit code + URL only

# Fully offline scoring — your prompts never leave your machine
rubric-chat rate <id> --local

Local mode (--local)

rate --local scores the session entirely on-device with a small open-weight model — no account, no API call, no monthly cap, nothing uploaded. The default is Gemma 4 E2B (Apache 2.0) running on ONNX Runtime via transformers.js. The first run downloads the model (~3.2 GB, cached in ~/.rubric/models/); after that it's offline forever.

rubric-chat rate <id> --local              # default (gemma-4-e2b, ~5 GB RAM)
rubric-chat rate <id> --local --model gemma-4-e4b  # bigger judge (~8 GB RAM)
rubric-chat rate <id> --local --model gemma-3-1b   # low-RAM machines (~2 GB)
rubric-chat rate <id> --local --yes        # skip the first-run download prompt

Available models: gemma-4-e2b (default), gemma-4-e4b, gemma-3-4b, gemma-3-1b. Gemma 4 runs on the ONNX engine; Gemma 3 runs on llama.cpp with grammar-enforced output.

Local scores use the exact same rubric, weights, and strictness curve as the server, but a small on-device judge is not a frontier judge — expect roughly ±10 points (local tends generous) and plainer feedback. The report is badged LOCAL (approximate); run without --local for the calibrated score and a shareable card.

Supported sources

| Source | How it's discovered | Status | | --- | --- | --- | | Claude Code | ~/.claude/projects/**/*.jsonl | ✅ | | Codex CLI | ~/.codex/sessions/**/*.jsonl | ✅ | | Cursor | local SQLite at ~/Library/Application Support/Cursor/User/globalStorage/state.vscdb (macOS) — verified against composerData / bubbleId schema _v: 16 | ✅ | | Cline / Roo / Aider / Continue | per-tool storage | ⏳ planned | | ChatGPT export | upload conversations.json via --file | ✅ | | Claude.ai export | upload Anthropic data export JSON via --file | ✅ |

What gets graded

The whole session as one artifact, on six weighted dimensions, with a strict curve so a competent median lands near 50 (not 85). Full rubric, weights, and the strictness formula are at https://rubric.chat/methodology.

What we look at

Only your turns. The CLI strips assistant text locally before any payload leaves your machine — see the apps/cli/src/sources/*.test.ts files in the repo, which assert this at the wire level. User turns are also PII-redacted on the server before they reach the scoring model.

Configuration

| Env var | Default | What it does | | --- | --- | --- | | RUBRIC_API_BASE_URL | https://api.rubric.chat | Override for local dev | | RUBRIC_WEB_BASE_URL | https://rubric.chat | Override for local dev | | RUBRIC_LOCAL_MODEL_URI | registry default | Override the --local GGUF model URI |

Credentials are stored in the OS keychain (macOS Keychain, libsecret on Linux). Fallback: ~/.rubric/credentials (chmod 600).

License

MIT