Repository navigation
Releases
ModelPick 0.2.0
Text execution adapters (OpenAI, Anthropic, Gemini, DeepSeek, OpenRouter) plus MCP run_model and local Ollama.
ModelPick 0.1.5
Keyless local Ollama run path; 205 tests, 90.18% coverage.
ModelPick 0.1.4
Local key store, CLI wizard, localhost key page, doctor status, and keyless Ollama detection.
ModelPick 0.1.3
Adds judge catalog for Jev and Gemini Decision 1; 182 tests, 90.10% coverage.
ModelPick 0.1.2
MCP server reports the package version; 180 tests, 90.02% coverage.
ModelPick 0.1.1
Official npm package plus MCP Registry metadata. 179 tests, 90.02% coverage, 28/28 adversarial.
ModelPick 0.1.0 — decision foundation
ModelPick 0.1.0 — decision foundation
The first public cut of a model router that prices cost per solved task, not per token.
What ships
-
40 models · 20 task nodes · 18 providers across coding, image, video, 3D, and audio.
-
pickreturns one recommendation with quality, expected attempts, cost per solved task, latency, why, trade-offs, and the exact key/env setup. -
Key-aware fallbacks. If you do not own the recommended provider,
picks.availablegives the best runnable model.--availablefilters to keys already present. No keys is not an error. -
Value, defined. The value pick is the cheapest model within 8 quality points of the best eligible model; if the band is empty, it falls back to the top pick. It keeps 95.4% of frontier first-pass quality at 41.6% of the cost in the simulation. Definition, simulation, and sensitivity:
docs/VALUE.md. -
7 MCP tools over stdio:
pick_model,get_decision_packet,plan_workflow,compare_models,get_setup,list_models,explain_task. -
Embedded everywhere it can be: Claude Code plugin + marketplace, Codex MCP/skill, Cursor, DeepSeek Harness profile row, Copilot, generic MCP.
-
Bring your own decision model.
modelpick decideemits a portable packet;modelpick pick --verdictapplies a judge/decision-model ranking locally. The verdict can only reorder the shortlist by two positions and cannot override constraints.pick_modelaccepts the same object over MCP. Protocol:docs/JUDGES.md. -
Evidence: 174 tests, 90.02% line coverage, 28/28 adversarial cases, 115 dev + 35 held-out decision runs, and an explicit confidence label on every quality score and price.
-
93-second intro video with a local English voice, natural English + 中文 captions, script, storyboard, and poster.
-
Independent image correction: Nano Banana Pro was lowered from 96 to 91 after Artificial Analysis ranked it #18, below Nano Banana 2.1 (#8) and GPT Image 2 high (#6); Veo 3.1 was lowered from 95 to 89 after the video arena ranked it #26, below Seedance 2.5 (#4).
docs/MEDIA_EVIDENCE.md.
Honest limits
- Media quality scores are priors, not independent measurements.
- The decision benchmark labels are expert-authored; it is not a live output benchmark.
- The retry curve is a model, not a measured correlation.
- The npm package is not published yet; install is
npx -y github:Leizhidong-creator/modelpick. - CI workflows are in
docs/ci/because the publishing token lacks the GitHubworkflowscope.
Install
npx -y github:Leizhidong-creator/modelpick pick "refactor this repo"
npx -y github:Leizhidong-creator/modelpick setup claude-codeMIT. No telemetry. No key storage. No gateway.