ModelPicker MCP
Nanman5/modelpicker
Fonte de verdade neutra sobre LLMs para agentes escolherem modelos com dados reais
Ver no GitHubServidor MCP que fornece a agentes de IA fatos atualizados e normalizados sobre o mercado de LLMs, sem opiniões. Inclui dados sobre modelos, custos, capacidades, benchmarks e provedores com rastreamento de fontes e datas de coleta.
Instalação
npx convez-ai Nanman5/modelpicker1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Agentes de IA selecionarem modelos com base em critérios objetivos (custo, contexto, benchmarks)
- Comparar capacidades e preços de LLMs de diferentes provedores
- Manter decisões de modelo atualizadas com dados frescos e verificáveis
Veredito

Ferramenta essencial para construir sistemas que escolhem modelos de forma fundamentada. Remove alucinações sobre preços e contexto ao fornecer dados estruturados com rastreamento de origem — ideal para aplicações que precisam tomar decisões de modelo programaticamente sem viés de provedor.
README
A tradução para o português deste README ainda está na fila.
modelpicker
Fresh, normalized facts about LLMs — for the AI agents that have to pick one.
modelpicker is an MCP server that gives Claude, Codex, Cursor, Cline, and any other MCP-aware agent a clean source of truth about the current LLM landscape: which models exist, what they cost, what they can do, and how they score on independent benchmarks.
It is deliberately opinion-free. There is no recommend_model tool, no "winner" in compare_raw, no aggregated ranking. The agent gets raw, sourced facts and decides for itself.
Why
When you ask an agent to pick a model:
- Its training data is months out of date.
- It has provider bias (Claude likes Claude, GPT likes GPT).
- It hallucinates prices, context windows, and capabilities.
- Or it just defaults to whatever was popular a year ago.
modelpicker removes the guessing. Every fact carries the URL it came from and the date it was scraped.
Installation
Add it to your MCP-compatible client. The dataset is bundled with the package, so it works offline.
For AI agents: self-install prompt
Paste this into any agent that can edit files and run shell commands (Claude Code, Cursor, Codex, etc.):
Install the MCP server at
https://github.com/Nanman5/modelpicker. Add an entry undermcpServers.modelpickerin my MCP config (Claude Code:claude mcp add modelpicker -- npx -y github:Nanman5/modelpicker; Claude Desktop:~/Library/Application Support/Claude/claude_desktop_config.json; Cursor:.cursor/mcp.jsonor~/.cursor/mcp.json; Codex:~/.codex/mcp_servers.json). Usecommand: "npx"andargs: ["-y", "github:Nanman5/modelpicker"]. After installing, restart the client and call thelist_providerstool to verify it connected.
Manual install — Claude Code
claude mcp add modelpicker -- npx -y github:Nanman5/modelpicker
Manual install — config file (Claude Desktop / Cursor / Cline / generic .mcp.json)
{
"mcpServers": {
"modelpicker": {
"command": "npx",
"args": ["-y", "github:Nanman5/modelpicker"]
}
}
}
The first launch clones the repo and builds via the prepare script (~10 seconds). Subsequent launches are instant.
Smoke test (no MCP client needed)
echo '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"smoke","version":"0"}}}
{"jsonrpc":"2.0","method":"notifications/initialized","params":{}}
{"jsonrpc":"2.0","id":2,"method":"tools/list","params":{}}' | npx -y github:Nanman5/modelpicker
Should print one JSON-RPC response per request. If you see the six tools listed in the second response, you're good.
When modelpicker-mcp is published to npm, replace github:Nanman5/modelpicker with modelpicker-mcp.
What the agent gets
Six tools. Each response includes dataset_generated_at so the agent can reason about freshness.
| Tool | Purpose |
|---|---|
list_models | List models, optionally filtered by provider, capability, modality, max input price, or min context window. Excludes deprecated by default. |
get_model | Full record for one model id (or alias). |
list_providers | All providers in the dataset with model counts. |
get_provider | One provider with all of its models embedded. |
list_benchmarks | Raw benchmark scores, optionally filtered by model, benchmark name, or source. No rankings, no aggregates. |
compare_raw | Side-by-side matrix of two or more models. No "winner". |
Example get_model output:
{
"dataset_generated_at": "2026-05-07T06:17:00.000Z",
"model": {
"id": "anthropic/claude-opus-4-5",
"provider": "anthropic",
"name": "Claude Opus 4.5",
"aliases": ["claude-opus-latest"],
"context_window": 1000000,
"max_output_tokens": 64000,
"pricing": {
"input_per_million_usd": 15,
"output_per_million_usd": 75,
"cache_read_per_million_usd": 1.5,
"cache_write_per_million_usd": 18.75
},
"modalities": { "input": ["text", "image"], "output": ["text"] },
"capabilities": ["tool_use", "vision", "extended_thinking", "prompt_caching"],
"knowledge_cutoff": "2026-01",
"released_at": "2025-10-15",
"deprecated": false,
"deprecation_date": null,
"benchmarks": [
{
"name": "mmlu-pro",
"score": 84.2,
"unit": "%",
"source": "https://artificialanalysis.ai/models",
"source_name": "Artificial Analysis",
"as_of": "2026-05-07"
}
],
"last_updated": "2026-05-07",
"sources": [
{
"type": "pricing_page",
"url": "https://www.anthropic.com/pricing",
"scraped_at": "2026-05-07T06:17:00.000Z"
}
]
}
}
V1 coverage
Providers: Anthropic, OpenAI, Google, xAI. Benchmark sources: Artificial Analysis, Aider Polyglot leaderboard.
More providers (Mistral, DeepSeek, Meta, Cohere) and more benchmark sources (LMArena, Vellum) are straightforward to add — see Contributing below.
How the dataset stays fresh
- A GitHub Action runs daily (and on-demand) and executes the scraper pipeline.
- Each provider has a per-file scraper in
src/scrapers/. Scrapers fetch HTML, clean it, then use an LLM (NVIDIA Build API by default) to extract structured records — robust against page-layout changes. - A separate stage runs benchmark scrapers (
src/scrapers/benchmarks/) and attaches scores to models by id / alias / name, never silently misattributing. - The pipeline writes
data/models.jsonand commits it. The full history lives ingit log -p data/models.json. - The published npm package bundles the latest committed dataset, so MCP clients work offline.
Sanitization & integrity
The page → LLM → dataset pipeline has several guards against bad data:
- HTML cleanup —
<script>,<style>, navigation chrome, hydration placeholders, inlinestyle=andon*=attributes are stripped before text extraction. Bare URLs and zero-width / bidi characters are removed (URLs are noise; the real source URLs are tracked separately in each model'ssourcesarray). - Prompt-injection guard — page content is wrapped in
<page url="..." type="...">…</page>and the system prompt explicitly instructs the LLM to treat anything inside those tags as data, never instructions. Literal</page>substrings in the page are escaped so a hostile page can't close the tag early. - Anti-hallucination — the system prompt forbids prior-knowledge fill-ins ("
claude-1", "gpt-3.5"). The LLM must emit only models whose names appear verbatim in the page text. - Cross-provider guard — marketplace pages (Vertex AI, Bedrock, Azure) list third-party models hosted on the platform. A name-pattern check drops models whose slug clearly belongs to another provider (e.g. a
claude-*model returned by the Google scraper is rejected and logged). - Per-row Zod validation — every extracted row is validated against the strict
ModelFactschema. Invalid rows are dropped with a warning; the rest of the batch survives. - Pipeline tolerance — if a single page fetch, single scraper, or single benchmark source fails, the rest still run and the dataset retains prior data for the failing piece.
Together these mean a malicious or misformatted page can't poison the dataset: the worst case is "no fresh data for that provider this run, prior data kept", not "wrong data committed".
Local development
git clone <this repo>
cd modelpicker
npm install
cp .env.example .env # add your NVIDIA API key (free at https://build.nvidia.com)
npm run typecheck
npm test
# Run the scraper pipeline locally:
npm run scrape -- --dry-run # see what would be written
npm run scrape -- --only anthropic # one provider
npm run scrape -- --skip-benchmarks # provider data only
npm run scrape -- --benchmarks-only # refresh benchmark scores only
# Run the MCP server in dev:
npm run dev
# Build for publish:
npm run build
Environment variables
| Variable | Required | Default |
|---|---|---|
NVIDIA_API_KEY | yes (pipeline only) | — |
NVIDIA_EXTRACTION_MODEL | no | meta/llama-3.3-70b-instruct |
NVIDIA_BASE_URL | no | https://integrate.api.nvidia.com/v1 |
MODELPICKER_DATA | no | bundled data/models.json |
The MCP server itself never calls the LLM — only the pipeline does. End users do not need an API key.
Contributing a new provider
- Add a file
src/scrapers/<id>.tsthat exports aScraperconstant: provider metadata, the URLs to fetch, and any extraction hints. - Register it in
src/scrapers/index.ts. - Add a fixture-based test in
tests/scrapers/(mock the LLM and assert the run loop attaches sources correctly).
Same idea for benchmark sources under src/scrapers/benchmarks/.
Architecture
┌──────────────────┐ ┌──────────────────┐ ┌──────────────────┐
│ Scraper pipe │────▶│ data/models. │────▶│ MCP server │
│ (GH Action) │ │ json (committed)│ │ (npm package) │
└──────────────────┘ └──────────────────┘ └──────────────────┘
daily / manual single source consumed by
of truth, in git Claude/Codex/etc
Full design rationale: docs/superpowers/specs/2026-05-07-modelpicker-design.md.
License
MIT