Overview
AI API Pricing publishes its whole catalogue as an MCP server, so an assistant like Claude can look up model prices, compare options, and check what changed — without you copying numbers out of a browser tab.
- URL:
https://aiapipricing.slxd.app/mcp/aiapipricing - Authentication: none. The connector is public and read-only — every call is served the same confidence-scored catalogue anyone can read over plain HTTP, so there is no credential to configure.
- Underlying data: every tool is a thin wrapper over the same query layer the REST API uses. A tool never computes its own answer — the one bug this project cannot afford is an assistant quoting a different price than the website.
Connecting
Claude, ChatGPT, or another connector-based client: add
https://aiapipricing.slxd.app/mcp/aiapipricing as a connector. No sign-in step —
anonymous access is granted immediately.
Anything that speaks MCP directly (agent frameworks, scripts, the MCP Inspector):
npx @modelcontextprotocol/inspector
# URL: https://aiapipricing.slxd.app/mcp/aiapipricingNo header or token is required.
Tools
All seven tools are read-only (readOnlyHint: true) — nothing they do is
metered, gated, or written to an audit trail beyond ordinary rate limiting.
search_models
Searches the catalogue and returns current prices, confidence, and sources.
| Argument | Type | Default | Notes |
|---|---|---|---|
query | string | — | Free text matched against name and id |
provider | string | — | Restrict to one provider slug, e.g. openai |
limit | number | 20 | 1–100 |
get_model_price
Every current price for one model — input, output, cached reads, batch and
priority tiers, context brackets — with confidence and which sources agree.
Two dimensions can share a label, so the ones that would collide carry their
qualifiers, e.g. Web search (context size: high).
| Argument | Type | Notes |
|---|---|---|
model | string | provider/slug, e.g. anthropic/claude-sonnet-4-5 |
channel | string | Sales channel to price. Omit for the maker's own price: direct, vertex, bedrock, azure, alibaba, openrouter, coding_plan, token_plan, other. A channel that does not sell this model returns no prices |
compare_models
Prices of several models side by side.
| Argument | Type | Notes |
|---|---|---|
models | string[] | 2–8 model ids, provider/slug |
Amounts billed in credits are never converted to money, so a credit-billed
model is not directly comparable to a token-billed one. Two dimensions can
share a label, so the ones that would collide carry their qualifiers, e.g.
Web search (context size: high).
price_history
A model's price over time — one point per change, not per day, since prices
are step functions. One series per price dimension: two dimensions can share a
label, so the ones that would collide carry their qualifiers, e.g.
Web search (context size: high).
Each point is [date, amount], the date as YYYY-MM-DD; a price that changed
twice on one day comes back as two points with that same date, in order.
| Argument | Type | Notes |
|---|---|---|
model | string | provider/slug |
from | string | ISO date, inclusive, optional |
to | string | ISO date, inclusive, optional |
recent_changes
Recent price increases, decreases, additions and removals across the catalogue, newest first.
| Argument | Type | Default | Notes |
|---|---|---|---|
days | number | 30 | 1–365 |
limit | number | 50 | 1–200 |
list_providers
Every provider in the catalogue and how many models each has. No arguments.
These are the houses that make the models: a reseller (Vertex, Bedrock, Azure, OpenRouter, Alibaba's catalogue, a subscription plan) is a sales channel, not a provider — ask get_model_price with a channel for what one of those charges.
data_sources
Where prices come from: which sources are consulted, their weight and licence, when each last succeeded, and whether any is currently failing. Use this to judge how much to trust a price. No arguments.
Safety and limits
- Rate limits: 120 calls/minute per credential and 240/minute per IP, checked before any lookup runs.
- Body cap: 1 MiB per request.
- Every tool is read-only, so there is no confirmation gate and no write path to guard — the connector cannot change anything in the catalogue.
See also
API reference documents the same data as plain REST, including the response envelope and caching headers, for callers that don't speak MCP.