llm-oracle

io.github.weiseerPublished by GitHub account weiseerv0.2.0Updated Oct 24, 2026

Up-to-date LLM pricing + availability catalog with cited sources for AI agent routing.

VerifiedSTDIODesktop onlyData & AnalyticsAI & ML

Overview

AI-generated overview

A local MCP server that gives an assistant a cited catalog of LLM provider pricing and availability, so it can compare models and find the cheapest option for a

What you can do
List models, optionally filtered by provider or capability, and fetch the full record for one model with its source links and timestamps. Rank models by estimated cost for a given number of input and output tokens, optionally requiring certain capabilities. Compare several models side by side and check a model's current availability status with its cited source. The catalog covers five providers and about ten models in this version.
When to use it
Useful when an assistant has to choose between LLM providers or models at request time and needs current prices rather than a hardcoded table. Also handy for auditing or quoting where a pricing claim comes from, since each row links to the provider's documentation.
Requirements
Runs as a local process; the README's setup uses Python with a cloned repository, while the manifest lists an npm package. Network access to the hosted catalog by default; set LLM_ORACLE_LOCAL_ONLY to use only the bundled catalog, or LLM_ORACLE_URL to point at another catalog. No account or key is needed for the free tier.
Before you install
The self-hosted MCP server logs nothing, but the hosted catalog service records each request (caller, query summary, latency, paid status) in an audit ledger. Paid tiers cost USDC per month and are arranged by email; the free tier is limited to 1,000 calls per day per IP. Pricing data can go stale, so check the as_of timestamp and cited sources before relying on it.

Installation

In SourceWeft

  1. Open llm-oracle in the dashboard and add it to a workspace.
  2. 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

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.)

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

Add to your MCP client config:

json
{  "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
  • get_model(model_id) → full record for one model
  • find_cheapest(input_tokens, output_tokens, required_capabilities?) → ranked by estimated cost
  • compare_models(model_ids[], input_tokens?, output_tokens?) → side-by-side
  • check_availability(model_id) → current status with cited source

As an HTTP API

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

As raw JSON

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

Pricing

TierCallsCost
Free1,000/day$0
Pro100,000/month$5 USDC/mo
Scale1,000,000/month$20 USDC/mo
Raw catalogunlimited$0 (always free)

Paid tiers settle in USDC. Email [email protected] 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 model_id, provider, family, context_window, max_output_tokens, input_price, output_price, cached_input_price, 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.

Roadmap (probe-pulled, not pre-planned)

This is P-001 in weiseer's strategic backlog — a market sensor for organism's architecture v2. We expand it where users ask, kill it if 4 weeks of public availability brings under 100 calls. Don't expect a polished commercial product; expect honest experimentation with daily updates.

What we might add (in observed-demand order):

  • More providers (Cohere, xAI, AI21, Together, Replicate hosted, Bedrock, Azure pricing variants)
  • x402 micropayments for high-volume agent buyers (deferred — see arXiv 2605.11781 on protocol security)
  • Historical pricing time-series
  • SLA + uptime aggregates per provider
  • Per-region availability where providers vary
  • Self-hosted model cost estimators (with caveats)

Contact

[email protected] · github.com/weiseer · weiseer.com

Source: README.md at commit d45f6f1

Tools

0
SourceWeft has not started this package in its sandbox yet, so its tools are not known.

Version history

1
  1. v0.2.0LatestSep 16, 2026

Score

40% 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

    Declare a license in the repository or the registry entry.