Find the best LLM for your codebase—not someone else’s benchmark.

ModelFit runs repo-specific coding probes across candidate models, grades their answers blindly against explicit rubrics, and ranks correctness before cost and latency. Public benchmarks measure average code; ModelFit asks whether a cheaper or secondary model can handle your SwiftUI, your Drizzle migrations, your Cloudflare Worker, and your failure modes.

target repo ──▶ probes (PROMPT + RUBRIC) ──▶ run.sh ──▶ candidate answers
                                                              │
        attempts.csv + verdicts.csv ◀── judge.sh ◀────────────┘
                    │
                report.sh ──▶ coverage-aware leaderboard

Why it is different

• Your workflow, not a generic suite. Probes are generated from a target repo you name explicitly.
• Any compatible model. OpenAI-compatible /chat/completions and Anthropic-compatible /v1/messages endpoints.
• Blind rubric grading. The judge sees the task, rubric and answer, not the candidate model name.
• Correctness first. Cost and latency never rescue a correctness loss.
• Auditable runs. Every run gets an immutable run ID, per-sample outputs, attempt ledger and verdict ledger.

Security and data boundary

ModelFit is designed so secrets and run outputs are excluded from Git by default, but no local tool can guarantee you will never leak sensitive data.

• config/models.json stores only the environment variable names that hold keys. The real keys live in your shell or .env, which is gitignored.
• .env, config/models.json, runs/ and results.csv are ignored.
• bin/scan-secrets.sh checks tracked files for common secret-shaped strings before publishing.
• Generated probes may contain proprietary code, customer data, credentials or personal data. Review probes before running them.
• Probe prompts are sent to each configured candidate provider. Task, rubric and candidate answer are sent to the judge provider.

Quickstart

git clone https://github.com/kwadwoadu/modelfit.git
cd modelfit
brew install jq shellcheck   # shellcheck optional, for local linting

./bin/selftest.sh            # zero API spend; includes mock-provider tests

cp config/models.example.json config/models.json   # edit models + judge
cp .env.example .env                                # paste keys; never commit
./bin/modelfit doctor --repo ../your-app

Generate probes with Claude Code from the ModelFit repo:

/modelfit --repo ../your-app

Then smoke-test one probe/model before the full suite:

./bin/modelfit run example-chunk fake-model-key --samples 1
./bin/modelfit judge example-chunk fake-model-key
./bin/modelfit report

Full run:

for p in probes/*.md; do
  n=$(basename "$p" .md)
  ./bin/modelfit run "$n" all --samples 1
  ./bin/modelfit judge "$n" all
done
./bin/modelfit report

If one model fails, the batch continues where possible but exits non-zero and the report shows incomplete coverage.

Add your workflow

1. Agent-generated probes. Run /modelfit --repo ../your-app. The command inspects the target repository, writes 6–10 probes into probes/, and records non-sensitive provenance.
2. Manual probes. Copy probes/example-*.md: a # PROMPT sent to each model and a # RUBRIC the judge grades against.

A good probe has one decisive discriminator: the subtle thing a weaker model gets wrong.

How scoring works

• run.sh sends each probe to candidates, strips markdown fences, retries empty/truncated replies up to the token ceiling, and records every attempt in runs/<run-id>/attempts.csv.
• judge.sh sends task + rubric + untrusted candidate answer to the judge, validates strict JSON verdicts, and writes runs/<run-id>/verdicts.csv.
• report.sh ranks by pass percentage, quality and candidate cost, while showing judged count, attempts, incomplete attempts and actual recorded total cost. Add --by-task for a per-probe candidate-cost breakdown (which kinds of task are expensive on which model).
• Candidate cost, judge cost and retry cost are tracked from provider token usage when available. Missing usage is NA, not zero.

Limitations

• LLM judges are useful but not objective. Blind labels reduce model-identity bias; they do not remove style bias or prompt-injection risk.
• Judge-only probes do not execute candidate code. If compilation is decisive, add an executable gate in a future probe.
• Prices in config/models.example.json are placeholders. Verify provider pricing before trusting cost comparisons.
• One sample is not statistical confidence. Use --samples N when run-to-run variance matters.
• Provider “compatibility” varies. Use ./bin/modelfit doctor and a smoke probe before a large run.

Layout

modelfit/
├─ bin/    modelfit run.sh judge.sh report.sh doctor.sh selftest.sh scan-secrets.sh
├─ bin/lib/common.sh
├─ config/ models.example.json
├─ probes/ example-honesty.md example-chunk.md
├─ prompts/ generate-probes.md judge-system.md
├─ tests/ mock-provider reliability tests
├─ .claude/commands/modelfit.md
├─ examples/ results.example.csv .env.example .gitignore LICENSE

MIT licensed. Built by Kwadwo Adu.