Hosted audit API
Find out what your agent run wasted
Send a finished run's metrics and get back the share of tokens that did not need to be spent, the reason, and an approximate dollar figure with the price snapshot it came from.
Endpoint
POST https://api.mcp-share.com/v1/audit Authorization: Bearer <your key>
Also available as an Apify Actor, which handles the call and billing for you. A native MCP server for this API is planned.
Why it runs after the run
An agent cannot catch this about itself
The most common finding is a prompt-cache prefix that broke partway through — a tool added, a memory file edited — after which a prefix that was being served from cache gets billed as fresh input. That pattern only exists when you compare consecutive calls across the whole trace. An agent in the middle of working has no view of what it is about to do, and asking it to check would itself cost tokens.
What it reads
Token counts, model ids, cache hit ratios, timestamps, latency, event types.
Prompts, replies, documents, tool payloads, memory contents, system prompts.
Traces with no model calls. An empty trace is refused rather than charged.
What it finds today
One analyzer is in production: cache.unstable-prefix, which
detects the point where a cached prefix stopped matching and quantifies
what that cost. Baseline token accounting and cost coverage apply to every
trace regardless. More analyzers are on the roadmap; this page will say so
when they ship, not before.
Pricing
By trace size| Tier | Trace size | Price |
|---|---|---|
| Estimate | any | free |
| Small | up to 100k tokens | $0.05 |
| Medium | 100k – 1M | $0.19 |
| Large | 1M – 10M | $1.49 |
| X-large | over 10M | $4.99 |
Tokens are what we measure; dollar amounts are derived from a dated price snapshot and always carry their coverage. A model with no rate on file still produces a full token result — it just reports no dollar figure rather than inventing one.