llm-oracle MCP server

llm-oracle

Continuously-updated catalog + query API for LLM provider availability and pricing.

Built by weiseer.

What it does

Given a question like "for this prompt of N input tokens plus M expected output tokens, which currently-available model offers the lowest cost?" the oracle answers in milliseconds, with the source URL it used for the pricing claim.

Use it as:

• an MCP server for AI agents that need to make routing decisions at inference time
• an HTTP JSON API for any tool that wants up-to-date LLM pricing/availability
• a raw catalog.json you can audit, vendor, or contribute fixes to

Why

LLM pricing changes monthly. Every multi-model app rebuilds the same lookup table. We maintain it so you don't.

• Cited: every pricing row links to the provider's official documentation
• Versioned: as_of timestamp on every response so you know how fresh the data is
• Open: catalog is public JSON, MIT licensed; client/server code is Apache-2.0
• Cross-checked: source data verified against the open-source LiteLLM price file where overlap exists

Coverage (v0)

5 providers, ~10 models. Growing as we observe demand.

• Anthropic: Claude Opus 4.7, Sonnet 4.6, Haiku 4.5
• OpenAI: GPT-5, GPT-4o
• Google: Gemini 2.5 Pro, Gemini 2.5 Flash
• DeepSeek: Reasoner, Chat
• Mistral: Large

Quickstart

As an MCP server (Claude Desktop, Cursor, Continue, Cline, etc.)

git clone https://github.com/weiseer/llm-oracle
cd llm-oracle

Add to your MCP client config:

{
  "mcpServers": {
    "llm-oracle": {
      "command": "python",
      "args": ["/absolute/path/to/llm-oracle/mcp_server.py"]
    }
  }
}

The agent now has tools:

• list_models(provider?, capability?) → list models, optionally filtered
• getmodel(modelid) → full record for one model
• findcheapest(inputtokens, outputtokens, requiredcapabilities?) → ranked by estimated cost
• comparemodels(modelids[], inputtokens?, outputtokens?) → side-by-side
• checkavailability(modelid) → current status with cited source

As an HTTP API

# Free tier: 1,000 calls/day per IP, no key
curl https://oracle.weiseer.com/catalog.json
curl 'https://oracle.weiseer.com/cheapest?input_tokens=2000&output_tokens=500'
curl https://oracle.weiseer.com/models/claude-sonnet-4-6

As raw JSON

curl https://oracle.weiseer.com/catalog.json > my-local-catalog.json

Pricing

Paid tiers settle in USDC. Email wei@weiseer.com to subscribe; we issue an API key and confirm receipt. (We'll automate this when the first three paid customers exist.)

Schema

See catalog.json for the canonical schema. Per-model fields include modelid, provider, family, contextwindow, maxoutputtokens, inputprice, outputprice, cachedinputprice, capabilities, availability_status, source URLs, and timestamps.

Update cadence

• Pricing: daily cron + manual when a price change is publicly announced
• Availability: 5-minute poll of provider status pages where available; daily smoke-test otherwise

Errors and disagreements

Found a stale price? An incorrect capability flag? A missing model? Open an issue — please include the source URL we should be tracking.

Telemetry

The hosted service logs each request as a single event (caller, query summary, latency, paid status) to weiseer's append-only audit ledger. The MCP server you self-host logs nothing.

Building on this

The catalog format is MIT licensed. Vendor it, mirror it, build on top of it. We ask only that you cite weiseer/llm-oracle somewhere visible if you redistribute the data.

The server and MCP code are Apache-2.0.

Shortened. The full README is on GitHub.