LiveLLM

Inspiration

I tried to find this as an existing product and was suprised when I couldn't. I wanted to find and compare the cheapest current deals on LLM inference. LiteLLM and openrouter had token pricing but they didn't analyse promotions of each provider and normalise them.

What it does

It discovers real-time LLM pricing changes through SerpApi, verifies them against official sources, and serves compact, provenance-backed economic facts via a REST API and MCP server.

How we built it

The pipeline has five stages:

1. SerpApi Discovery — Google News Light detects price events, Google Light resolves official sources, Search Index Deep explores unknown providers. JSON Restrictor reduces payloads. Account API governs quota.
2. Official Source Fetch — Prefilter scoring selects the most authoritative source for each candidate.
3. AI Structured Extraction — mimo-v2.5 extracts pricing facts from official pages: input/output/cached prices, context windows, rate limits, free tiers, promotions.
4. Deterministic Validation — 6 checks catch hallucinated or incorrect extractions: evidence quote matching, numeric type check, range validation, unit compatibility, confidence threshold (≥ 0.5), entity validity.
5. Fact Ledger — Bitemporal (valid_from/valid_to), provenance-backed, content-addressed (SHA-256). Every fact traces back: fact → evidence → source observation → official source URL → SerpApi search metadata ID. The REST API serves 7 endpoints. The MCP server exposes 5 read-only tools. A quota governor progressively tightens when SerpApi credits run low.

Challenges we ran into

Hallucination. Lots of it..

Accomplishments that we're proud of

Finishing a project and one that I believe is actually useful and I am actively using as infra in my other projects

What we learned

Its easy to get 80% of the way there and much harder to finish fully.

What's next for LiveLLM

x402 machine-readable payments — let agents pay for premium market data via protocol-level micropayments Cross-model capability tracking — extend beyond pricing to track context windows, rate limits, and capability changes over time Agent cost simulator — given a workload profile, compute the optimal provider+model route across all known options Wider provider coverage — expand from 23 models to cover the full landscape including open-source inference providers

Built With

• account-api)
• cloudflare-workers
• google-light
• json-restrictor
• mcp
• mimo-v2.5
• node.js
• search-index-deep
• serpapi-(google-news-light
• sqlite
• typescript