
ModelPick
io.github.Leizhidong-creatorPublished by GitHub account Leizhidong-creatorv0.1.1Updated Nov 19, 2026
Pick the best-value model at every task node, with cost per solved task and the exact setup.
The publisher's own server: model.tools — Zero-Auth Agent Utilities
Overview
Local stdio MCP server that recommends which AI model to use for each task node, with cost per solved task and connection setup.
- What you can do
- Pick and compare models for a task, with quality, cost per solved task and alternatives
- Plan multi-step workflows with one model per step and explain how a task was classified
- Browse the model catalog and list the supported task nodes
- Get connection details for a model: env var, key URL, docs and API model id
- Return a decision packet for an external judge model, or re-rank a shortlist with a verdict
- When to use it
- Useful when an agent or developer wants a defensible choice of model per step across text, image, video, 3D and audio, rather than defaulting to one frontier model for everything, and wants the trade-off between quality and cost made explicit.
- Requirements
- Runs locally over stdio via npx; Node.js 20.11 or newer, with no runtime dependencies. No account, key or network access is needed for recommendations. Provider API keys are only needed to actually run a recommended model, and the server reports which keys are present without reading their values.
Installation
In SourceWeft
- Open ModelPick in the dashboard and add it to a workspace.
- Enable the server for the chats that should use its tools.
Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.
Other MCP clients
Follow the launch instructions in the repository.
README
[ModelPick — pick the right model at every step]
ModelPick
The model router for AI agents — with cost per solved task.
Coding, image, video, 3D, and audio. One CLI, one MCP server, every agent.
[release] [license] [node] [MCP] [tests] [adversarial] [dependencies]
Website · intro video (93s) · How value is defined · Ecosystem
Your agent can write code, call tools, and ship a feature. Then it picks one model for everything — a frontier model to classify a support ticket, to generate a barrel prop, to draft a thumbnail.
The result is predictable: 60–95% of the spend goes to work that did not need a frontier model, and the hard nodes still fail because the model was chosen for the wrong modality.
ModelPick is a router and cost calculator, not another gateway. Before each node, it answers three questions in one call:
- Which model? Ranked for this exact task, with hard capability filters (context, text-in-image, rigging, clip length, open weights).
- What will it really cost? Cost per solved task — list price × expected attempts, not $/M tokens.
- How do I connect it? The env var, key URL, docs, and API model id.
[ModelPick CLI recommending Tripo H3.1 for a 3D game asset]
[Watch the ModelPick intro video]
▶ Watch the intro video with English + 中文 captions · direct MP4 · script & storyboard · English + 中文 captions
Why this exists
Three things are broken in model selection today:
1. Token prices hide the real bill. A model that is 10× cheaper per token but needs 1.8× the attempts and 2× the output tokens is not 10× cheaper. ModelPick computes the denominator that matters: dollars per usable output.
2. A coding agent defaults to one model, even for creative nodes. A 3D pipeline is not a text problem. Text-in-image is not image generation. A 30-second video is not a 5-second clip. ModelPick models 20 task nodes across 5 modalities, with capability filters that reject the wrong model before scoring.
3. Every tool is a new integration. ModelPick ships as a zero-dependency CLI, a stdio MCP server, a portable SKILL.md, and a Claude Code plugin. Same decision engine everywhere; no gateway, no lock-in, no key storage.
In one line: other tools route text prompts. ModelPick routes work — coding, image, video, 3D, and audio — and prices each node per solved task.
How is this different from a gateway or a coding router?
Gateways are useful, and ModelPick can recommend one as the setup path. But a gateway cannot tell you that a 3D node should use Tripo instead of Meshy, or that a "headline text" image needs a model with a text-rendering score above 90. That choice is what this tool makes.
More: research/competitive-landscape.md (15 products, 10 features, 53 sources).
30-second quickstart
Official npm package: https://www.npmjs.com/package/modelpick
MCP Registry: https://registry.modelcontextprotocol.io/v0/servers?search=modelpick
Fallbacks if the registry is unavailable:
[ModelPick planning a product launch video with a model per step]
Works with the agents you already use
modelpick setup writes the files for the agent you use. It backs up existing config, keeps sibling MCP servers, and refuses to overwrite invalid JSON.
Claude Code
Restart Claude Code, run /mcp, and approve the project server. Then ask: pick a model for a 3D sword for Unity.
Plugin alternative, inside Claude Code:
OpenAI Codex CLI
setup codex writes AGENTS.md and .agents/skills/modelpick/. Codex reads MCP servers from ~/.codex/config.toml, so --global --write appends the row or you can use codex mcp add:
Cursor
It writes .cursor/mcp.json, .cursor/rules/modelpick.mdc, .cursor/skills/modelpick/, and AGENTS.md. Open Cursor → Settings → MCP and confirm modelpick is connected.
DeepSeek Harness
DSH has no .mcp.json. MCP servers are Cordis rows in the profile patch:
It writes the portable skill to ~/.dsh/skills/modelpick/, appends AGENTS.md, and adds one MCP row to ~/.dsh/profiles/desktop/cordis.patch.yml (backing the file up first).
Restart DSH after this. Then ask it:
Without --write, setup dsh prints the row instead of editing a live profile. The MCP server itself can be tested without DSH:
GitHub Copilot CLI
It writes .mcp.json, .github/skills/modelpick/, and AGENTS.md. Restart Copilot and ask for a model pick.
VS Code / Copilot agent mode
It writes .vscode/mcp.json under servers (VS Code's schema, not mcpServers), the portable skill, and AGENTS.md. In VS Code run MCP: List Servers, start modelpick, then ask in agent mode. Sibling servers are preserved.
Any MCP client
Tools exposed: pick_model, get_decision_packet, plan_workflow, compare_models, get_setup, list_models, explain_task. Recommendations work without any provider key; the key is only needed to run the chosen model.
Portable skill: .agents/skills/modelpick/SKILL.md is read by Codex, Cursor, Copilot CLI, Devin/Windsurf, and DeepSeek Harness. Claude Code gets its own copy. The skill tells the agent when to call pick_model and how to report the result.
The number that matters: cost per solved task
Worked example — 80K input / 8K output, 60% cache hit, agentic coding:
The cheap model is 53× cheaper and 23.5 quality points worse. The interesting answer is the third row: mid-quality agentic coding at 2.3% of the frontier cost. The right answer depends on the node — which is why the tool shows both sides instead of hiding the trade-off behind a token price.
Quality is a weighted mix of public benchmark signals (BenchLM overall/coding, CursorBench 4.0, capability-specific scores). Expected attempts is a transparent model, shown on every result — not a promise. If a number is estimated, the CLI prints ~.
What it covers
40 models · 18 providers · 20 task nodes · 5 modalities. Current as of 2026-11-19; every price carries a source and read date.
The catalog is a dated snapshot, not a live price feed. modelpick update refreshes from the repo; modelpick doctor warns when it is stale.
Tested decisions
[ModelPick decision benchmark: 115 runs, 100% top-1 acceptable, 100% hard constraints]
What this is: 39 expert-labeled scenarios expanded into balanced/cost/quality runs, plus a 12-scenario held-out set authored after the benchmark froze. Every run checks task detection, top-1/top-3 acceptability, hard constraints, cost regret, and stability under ±20% price perturbation.
What this is not: a live model-output benchmark. It does not call providers, and it does not prove quality labels are universally correct. Quality scores are derived from public leaderboards; expected attempts is a model. The full methodology, failures, and reproduction commands live in tests/TEST_PLAN.md, tests/results/, and bench/.
[Quality versus cost per solved task for coding models]
Full per-node tables: docs/LEADERBOARD.md.
Value, measured
node bench/value.js simulates 4,000 jobs per task across five task types, using catalog quality as first-pass success and the router's retry curve:
The value pick keeps 95.4% of frontier first-pass quality at 41.6% of the cost. The cheapest model saves 2.5 cost points but loses 1.9 quality points and fails a 90%-parity bar on two of five tasks. This is synthetic; the arithmetic and the definition are in docs/VALUE.md, and the evidence behind every quality score is in docs/QUALITY.md.
Multi-dimensional scorecard
tests/results/SCORECARD.md combines six dimensions from committed evidence:
[ModelPick six-dimension scorecard]
Correctness is the decision benchmark; it is not an independent output benchmark. Cost is regret plus the value simulation. Robustness is price perturbation, retry sensitivity, and determinism. Coverage is every task and model. Safety is the 25-case adversarial suite. Operability is p95 latency, modelpick verify, and line coverage (84.8%).
The decision engine
Presets: balanced (task weights), value (price-performance), quality, cost, speed, privacy. Budgets and quality floors are hard constraints, not suggestions.
The default is a value decision, not a bargain hunt. 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 recommendation. It is returned as picks.value and value, with the quality gap and the percentage of the best model's cost. The full definition, simulation, and sensitivity analysis are in docs/VALUE.md.
Every answer includes:
recommendation.slug,quality,costPerTask,expectedAttempts,latencySecwhy[]andtradeoffs[]setup.direct(env, key URL, docs, API model) and an OpenRouter fallback for text modelscatalog.version,stale,priceConfidence
Bring your own decision model
The deterministic router is the default. If you want a judge or decision model (Claude, GPT, Gemini, a Google decision model, or a local model) to re-rank the shortlist, keep it outside the key store:
judges includes Jev (System One) (jev-latest, Venice, $0.042/1M input) and Gemini Decision 1 (gemini-decision-1, $0.10/1M input, zero output). They are decision models, not generation models.
The verdict can only reorder the shortlist, at most two positions, and cannot
invent a model or override a hard constraint. pick_model accepts the same
verdict object over MCP. The protocol, guardrails, and meta-evaluation
thresholds are in docs/JUDGES.md.
Commands
Useful flags: --optimize, --max-cost, --min-quality, --open-weights, --context, --exclude, --task, --verdict, --json, --explain, --no-color.
Only have a few API keys?
You do not need every provider. ModelPick separates the best decision from what you can run right now:
--availablerestricts ranking to models whose key exists (or any text model if you haveOPENROUTER_API_KEY).- The JSON always includes
availability.bestAvailable,availability.fallbacks, andrecommendation.availability.missing. - An
OPENROUTER_API_KEYis a universal fallback for text models;FAL_KEYcovers many media models. - If a capability has no accessible model, the tool says so instead of inventing a recommendation.
Add your own model
Bring a private endpoint, a self-hosted vLLM/Ollama model, or a provider the catalog does not cover:
The model is written to ~/.modelpick/overrides.json, validated against the full catalog before saving, and never touches the repo. You can also add a whole list with { "models": [...] }.
Connect without guessing the endpoint
pick gives you the env var and key URL. show gives you the full connection card:
The same object is returned by the MCP get_setup tool. If an endpoint is a provider's common path rather than a freshly verified one, the card says so (typical vs verified).
Built to change
Model prices move weekly. New providers appear monthly. The engine is deliberately data-first:
Custom model example:
No fork. No rebuild. modelpick doctor validates the result.
Data, honesty, and limits
- Prices: models.dev / OpenRouter for text; BenchLM media pricing, Google Vertex, OpenAI, fal, Tripo, Meshy, ElevenLabs for media. Read date on every source.
- Quality: BenchLM BenchAlign and CursorBench 4.0 for text; vendor specs plus public comparisons for media. Media quality is the least independently verified part of the catalog; it is labeled and dated.
- Known omissions: Veo is Vertex-only now; Sora 2 is discontinued; GPT-Image-1 and Imagen 4 are retired. These are deliberately not in the catalog.
- No telemetry. No key storage. Recommendations work offline.
keysreads presence only. - Not a gateway. ModelPick does not proxy requests, mark up tokens, or hold your data. You call the provider directly.
See docs/DATA.md for provenance and research/ for the underlying research.
Development
Requirements: Node ≥ 20.11. Runtime dependencies: zero. Python + Pillow are only needed to regenerate images.
- Architecture
- Data provenance
- How value is defined
- Quality evidence
- Media evidence: the first independent correction
- Bring your own eval
- Self-hosting
- NPM mirror
- Live endpoint evidence
- Ecosystem and embedding
- Judge/decision models
- Publishing and launch runbook
- Growth playbook
- Per-node leaderboard
- Model applicability
- Multi-dimensional scorecard
- Adversarial review
- Test form: 20 tasks × 40 models
- Citation metadata
- Design partner program: 5 slots
- Test plan
- Benchmark methodology
- Contributing
- Security
- Changelog
Roadmap
- Publish to npm (
npx modelpick) and submit to MCP directories - Live price refresh from provider APIs with signed snapshots
- Community-contributed task nodes and quality evidence
- Router mode: optionally execute the recommended call through the user's own key
- Cost ledger: track actual spend per node across a project
- Windows-native MCP shim (no
cmd /c)
If ModelPick saves you money, star the repo — it is the only distribution channel for a tool with no VC and no telemetry.
MIT © 2026 Lei Zhidong · built with zero runtime dependencies and a healthy distrust of token pricing.
Source: README.md at commit 6e5821d
Tools
0Version history
1- v0.1.1LatestNov 19, 2026
Score
50% of the applicable checks met
How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.
- Reached in the last 30 days25 pts
Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.
- Auth declared10 pts
Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.
- Tools listed15 pts
Answer tools/list without a credential, or ship a package that starts in a sandbox without one.
- DescriptionRequired15 pts
- Logo10 pts
- Documentation15 pts
- License10 pts
Other Model integrations

model.tools — Zero-Auth Agent Utilities
tools.modelOfficial publisher
Zero-auth remote utility server giving agents disposable email OTP reading, durable checkpoints, webhook bins, wake timers, receipts and page-to-Markdown…
NLP.COM.AI Open Model Catalog
ai.com.nlpOfficial publisher
NLP.COM.AI Open Model Catalog's hosted MCP server, published by nlp.com.ai; no sign-in required; Streamable HTTP; 3 tools.

Hologram Model Hub
foundation.uor.hubOfficial publisher
A remote MCP server that lets an assistant search an index of open models and get verified download links with SHA-256 checksums.

DOC Model Safety Gate
Published by GitHub account laceykes
A local MCP server that stages AI-proposed content changes for human review and only lets reviewed runs publish.

MCPShelf Model Context Protocol Registry, Search & Security Auditor
Published by GitHub account RitualDev-Lab
A local MCP registry browser that lets an assistant find, compare trust scores for, and get install snippets for 400+ MCP servers.