One click
With Cursor or VS Code installed, click a button: the app opens and asks you to confirm, and that is it.
In Claude (web or desktop): Settings → Connectors → Add custom connector, and paste the address below. Other tools are listed under Connect.
The address
https://auramatrix.ai/mcp
Streamable HTTP, no account or key needed. Ten tools only read; share_price alone stores something, and only after you explicitly agree. One request at a time per client, small responses, rate-limited.
Connect
Cursor
{
"mcpServers": {
"auramatrix": {
"url": "https://auramatrix.ai/mcp"
}
}
}One click opens Cursor and asks you to confirm. Or paste this into its MCP settings.
VS Code
{
"servers": {
"auramatrix": {
"type": "http",
"url": "https://auramatrix.ai/mcp"
}
}
}One click opens VS Code and asks you to confirm. Or put this in .vscode/mcp.json.
Claude
https://auramatrix.ai/mcpIn Claude: Settings, Connectors, Add custom connector, then paste this address.
Claude Code
claude mcp add --transport http auramatrix https://auramatrix.ai/mcpRun in a terminal, then ask Claude Code about token prices.
Codex
codex -c 'mcp_servers.auramatrix.url="https://auramatrix.ai/mcp"'Starts Codex with the index attached for this session.
Other clients
{
"mcpServers": {
"auramatrix": {
"url": "https://auramatrix.ai/mcp"
}
}
}Paste into the MCP settings of Windsurf or any client that takes a JSON config.
The eleven tools
list_models: every model with its official input and output price, lowest-priced channel, and page. Filter by vendor or text.get_price: one model across every channel: region, pricing mode, cached and batch rates, tiers, announced changes, trust status, observation time, and source snapshot.compare_channels: the same model across channels, sorted by input price, with the difference from the official list price.cheapest: models sorted by their lowest input price across all channels, with optional filters for vendor or channel type.price_changes: observed changes between records and vendor-announced changes with effective dates.estimate_cost: the cost of one request from token counts, applying the channel's long-context, time-of-day and thinking-token rules; a range and the reason where the source leaves a question open.gpu_prices: what renting a GPU costs per hour. Without arguments, every GPU model with its on-demand range across providers; with a GPU id, every provider, kind of price (on demand, spot, reserved, marketplace) and machine, with sources.price_index: the trial indices: token price indices for flagship, mid and small models, and the GPU-hour reference price per GPU model. For "are prices going up or down".measured_performance: what hourly test requests measured: time to first token, output speed, share of requests answered, and whether the charge matched the list price.check_price: places a price you pay or were quoted among the prices on record, for a model API or a GPU. For "is this a good price".share_price: adds a price, anonymously, to the benchmark of what buyers actually pay. Your assistant may call it only after you say yes; no account or address is stored, no single price is published, and a median appears once an item has five reports.
Things to ask
- "Where is Claude Opus 5.5 cheapest? Give me the source."
- "I send 20M input and 4M output tokens a day. What does that cost on three flagship models?"
- "Which models are cheaper than GPT-6.1 Sol and answer within a second?"
- "What does an H100 cost on demand at each provider right now?"
- "Which models changed price this week?"
Every answer ends with the same caveats as the pages: list prices as published at the observed time, the channel's own page is authoritative, aggregator prices are never the reference, and missing means unknown.
Without MCP
Plain HTTP, JSON back: https://auramatrix.ai/prices/v1/quote?model=openai/gpt-6.1-sol. Add &channel=azure-datazone-eu for one channel; model=list returns the ids. The whole dataset is at https://auramatrix.ai/prices/index.json, and a short guide for agents at https://auramatrix.ai/prices/v1/llms.txt.
What to expect
- Each record carries
trust_status: VERIFIED means an independent second reader re-read it from the source; UNVERIFIED means recorded, not yet re-read. - Field names will not change under you, but the HTTP endpoint is not yet a frozen API version. A
v1with history and change events is planned; see Methodology.