FiscaLLM MCP server: LLM prices and AI margins for agents

In short: add https://fiscallm.com/api/mcp to your AI assistant as a remote MCP server. It can then answer questions like "what would this feature cost on Claude Haiku instead of GPT-5 mini?" from a real price table, and, with your app's key, "which users cost me more than they pay?"

Connect

Claude Code:

claude mcp add --transport http fiscallm https://fiscallm.com/api/mcp

Cursor (.cursor/mcp.json):

{
  "mcpServers": {
    "fiscallm": { "url": "https://fiscallm.com/api/mcp" }
  }
}

VS Code (.vscode/mcp.json):

{
  "servers": {
    "fiscallm": { "type": "http", "url": "https://fiscallm.com/api/mcp" }
  }
}

In other clients, add a custom or remote MCP server with the URL https://fiscallm.com/api/mcp. The transport is Streamable HTTP and the server keeps no session.

Tools that need no key

Tool	What it answers
`list_llm_prices`	Current list prices of OpenAI, Anthropic and Google models, filtered by provider or name.
`get_llm_price`	One model's price, with its long-context rate and any announced change.
`estimate_llm_cost`	USD cost of a call from its token counts, and the total for a number of calls.
`compare_llm_costs`	The same call priced on several models, cheapest first.
`subscription_break_even`	How many messages a month a subscriber can send before AI cost uses up what they pay, after store or payment fees.
`fiscallm_setup_guide`	How to send usage (OpenTelemetry, LiteLLM, SDK or HTTP), connect revenue and read margins back.

Things to ask once it is connected:

"Compare the cost of 1,500 input and 400 output tokens on GPT-5 mini, Claude Haiku 4.5 and Gemini 2.5 Flash, for 300 calls a month."
"My plan is $9.99 a month through the App Store. How many messages can a user send on Claude Sonnet before I lose money?"
"How much does prompt caching save on this prompt?"

Tools for your own app

These read the same numbers as the dashboard. Send your app's ingest key as a header; keep it in an environment variable, not in a file you commit.

claude mcp add --transport http fiscallm https://fiscallm.com/api/mcp \
  --header "Authorization: Bearer $FISCALLM_KEY"

Tool	What it answers
`get_margin_overview`	AI cost, revenue, margin, AI's share of revenue and the daily series over a period.
`list_users_by_margin`	Each user's AI cost, revenue and margin, the ones losing the most money first.
`get_plan_margins`	AI cost, revenue and margin of each plan or product.
`get_cost_breakdown`	AI cost split by model and by feature.

Things to ask:

"Which users lost me money in the last 30 days, and on which feature?"
"What share of revenue went to AI this week compared with last week?"
"Which plan has the worst margin?"

No data yet? Create an app and send usage with OpenTelemetry, LiteLLM, the Vercel AI SDK or the SDK.

The same data over plain HTTP

Not using MCP? The price tools are also a free JSON API, and the margin tools are the read API. Both are described in openapi.json.

Details

Endpoint: POST https://fiscallm.com/api/mcp, JSON-RPC, one message per request.
Protocol versions: 2025-11-25, 2025-06-18, 2025-03-26 and 2024-11-05.
Read-only. No sessions, no server-sent events, no resources or prompts.
Open source (MIT): apps/api/src/mcp/route.ts. A self-hosted copy serves the same server at /api/mcp.

Questions

Is the MCP server free? Do I need an account?

The price and cost tools are free and need no account or key. The margin tools read your own app's data, so they need that app's key.

Can the server change anything in my account?

No. Every tool is read-only.

Where do the prices come from, and how fresh are they?

From FiscaLLM's open-source price table, checked against the OpenAI, Anthropic and Google pricing pages. Each answer includes the date the table was last checked.

Which clients work?

Any client that supports remote MCP servers over Streamable HTTP, including Claude Code, Claude, Cursor and VS Code. There is nothing to install or run locally.

Does the server see my prompts?

Only the arguments a tool is called with, such as a model name and token counts. FiscaLLM never stores prompts or responses from your app.