Skip to main content
Prometheus is an MCP server: AI assistants can connect to it and work directly with the platform’s structured, traceable financial data — statements, segments, filings, decks, transcripts, guidance, events, news, ownership, macro and the claim graph, across the US, the Gulf, Japan, Taiwan, Korea and Europe. Once connected, you don’t write API calls — you ask questions:
“Did NVIDIA beat its data-center guidance last quarter? Cite the filings.”
The assistant chains the tools itself (resolve the company → pull segment revenue → pull the guidance scorecard → trace the numbers to their filings) and answers with accession numbers next to every figure. There are 18 tools, all read-only, covering the live coverage inventory, search, standardized financials, fact-level citations, filings and their text sections, 8-K events, the news tape, earnings-call transcripts, guidance scorecards, point-in-time macro, and the causal claim graph. A connected app can never see your API keys, billing or account — those surfaces don’t exist on the MCP side.

Connect your client

No API key needed — you approve the connection with your Prometheus sign-in (OAuth).
1

Add the connector

In Claude, open Settings → Connectors → Add custom connector and paste:
2

Approve access

Claude sends you to Prometheus. Sign in (or you already are), review what’s being granted — read-only data access, metered on your plan — and click Approve.
3

Ask something

Back in Claude, enable the connector in the chat’s tools menu and ask a question. The connection works on Claude mobile too, and you can revoke it anytime from Claude’s connector settings.
Good first prompts: “What was Apple’s gross margin trend over the last 8 quarters? Cite each figure.” · “Summarize the newest 8-K from Tesla.” · “What did NVIDIA’s management promise on the last earnings call, and what’s their guidance hit rate?” · “Show US unemployment as it was known in March 2020.”

The tools

Text tools (get_transcript, get_event_text, get_section_text) return plain text with a [Source: …] citation line; everything else returns JSON identical to the corresponding REST endpoint.

Metering, auth and limits

  • A tool call is metered like the REST request(s) it makes — the same per-minute and monthly counters as the REST API. Most tools make one request; get_transcript (index + text) and get_guidance (commitments + scorecard) make two, and get_event_text makes two when it has to look the item code up first. Protocol plumbing (initialize, tools/list) is free. MCP calls appear in the usage dashboard as /api/mcp#<tool>.
  • get_financials takes the same period_type (including semiannual), reporting_basis and accounting basis as the REST route, and get_company returns the company’s accounting_bases so an assistant can ask for the right one.
  • Two credentials work: API keys (pk_live_..., as a Bearer or X-API-KEY header) and OAuth tokens from the connector flow. Both are data-scoped — key management, billing and account routes reject them by design.
  • When you’re rate limited, the tool result says how many seconds to wait; assistants honour it.

What the tools do not cover

The tool set is the read path an assistant needs most. Filing-note tables and series, the as-filed statement view, investor presentations, ownership, people, prices, benchmarks and correlation have no MCP tool today — the data catalog has the full map. Every one of them is a CLI command (prom note-tables, prom note-series, prom as-filed, prom presentations, prom ownership, …), and prom mcp runs alongside those commands, so an agent with the CLI installed reaches both.

Rules the server enforces on itself

The server’s own instructions carry the contract from agents.md: cite accession numbers next to figures, respect vintages (/financials is latest-restated and names its filing; macro is point-in-time), treat null as “not available” — never zero — and present data without investment advice.