Finance Pulse

Finance Pulse: Python client and MCP server

Financial news as structured, sourced data. Finance Pulse gives you financial news as statements: one-sentence facts from each article, with sentiment, importance, tickers, sector, source, headline and publication time. Statements about the same development are grouped into topics; one company's topics that continue each other form a storyline; and topics belong to standing themes. fintopic.news is a front page built on the same API.

The same setup guide, with the other guides and the API reference, is on the docs site: Use Finance Pulse from Claude, Cursor or Python.

This package gives you two ways to use it:

• a Python client for scripts, notebooks and pipelines;
• an MCP server, so Claude, Cursor and other MCP clients can query the news directly.

You need a RapidAPI key that is subscribed to Finance Pulse. Create a RapidAPI account, open the plans page and subscribe to a plan: there is a free one to try everything below. Then copy your X-RapidAPI-Key from the API's page on RapidAPI (the endpoint playground shows it). A key that is not subscribed to a plan is answered with 403.

Use it from Claude or Cursor (MCP)

The server runs with uvx, which is part of uv: install uv first. The package itself needs no separate install.

Claude Code

claude mcp add --env RAPIDAPI_KEY=YOUR_RAPIDAPI_KEY --transport stdio finance-pulse -- uvx finance-pulse

Claude Desktop and Cursor

• Claude Desktop: Settings > Developer > Edit Config opens claude_desktop_config.json (macOS ~/Library/Application Support/Claude/, Windows %APPDATA%\Claude\). Quit and restart Claude Desktop completely afterwards.
• Cursor: ~/.cursor/mcp.json

{
  "mcpServers": {
    "finance-pulse": {
      "command": "uvx",
      "args": ["finance-pulse"],
      "env": { "RAPIDAPI_KEY": "YOUR_RAPIDAPI_KEY" }
    }
  }
}

If the server does not start in Claude Desktop, it usually cannot find uvx: put the full path (the output of which uvx, or where uvx on Windows) in "command".

uvx keeps the version it installed first. To move to a new release, use finance-pulse@latest in place of finance-pulse once, or run uv cache clean finance-pulse.

Then ask things like:

• "What is going on with Nvidia? Give me the picture, then the key facts with sources."
• "Which tickers are suddenly getting more coverage than yesterday?"
• "What are the ongoing stories about Microsoft, and how did the biggest one unfold?"
• "What are the top developing stories in Energy, and how did the biggest one grow?"
• "Show me Tesla's daily news sentiment for the last two weeks."

Tools

Each tool call is normally one API request. Results are trimmed and come in small pages.

Use it from Python

pip install finance-pulse
export RAPIDAPI_KEY=YOUR_RAPIDAPI_KEY

The MCP SDK is installed with the package even if you only use the client.

from datetime import datetime, timedelta, timezone

from finance_pulse import FinancePulse

fp = FinancePulse()          # or FinancePulse("YOUR_RAPIDAPI_KEY")

# The newest key facts about a ticker
for s in fp.statements(symbol="NVDA", importance_min="high", limit=5)["data"]:
    print(s["published_at"][:16], s["sentiment"], s["statement"], f"({s['source_domain']})")

# One company in one call: counts per period, largest topics, storylines, newest statements
nvda = fp.symbol("NVDA")["data"]
week = nvda["periods"].get("w")           # a period without mentions is left out
if week:
    print(nvda["name"], week["mentions"], "mentions this week, score", week["score"])

# A company's news as continuing stories
for line in fp.storylines(symbol="MSFT", min_topics=2, sort="mentions", limit=5)["data"]:
    print(line["topics"], "topics,", line["mentions"], "mentions:", line["name"])

# Which tickers are suddenly in the news (kind= leaves out rates, central banks, countries, ...)
for row in fp.symbols(period="d", sort="change", kind=["equity", "etf"], limit=10)["data"]:
    print(row["id"], row["mentions"], row["change"])

# Daily news volume and sentiment, last two weeks
since = (datetime.now(timezone.utc) - timedelta(days=14)).strftime("%Y-%m-%dT00:00:00Z")
for b in fp.series(symbol="TSLA", interval="day", since=since)["data"]["buckets"]:
    print(b["t"][:10], b["count"], b["score"])

# What most outlets are reporting right now
for t in fp.trending(kind="live")["data"][:10]:
    print(t["sources"], "sources:", t["name"])

Every method returns the API's JSON unchanged: {"data": ..., "next_cursor": ..., "snapshot": {...}} (iter_statements and poll_feed yield its rows one by one). The snapshot block carries the times the data refers to (window_start, anchor, lag_days), and every statement carries published_at and indexed_at.

Follow the news without missing anything

poll_feed yields every new statement in the order it entered the API and keeps its position in a file, so it survives restarts: after a crash nothing is skipped and only the statement you were working on is delivered again. The first run, with no saved position, starts at the feed's default starting point (a limited time back, not the beginning of the data), and the position is first stored once that first page has been handled. After a long pause the poller reads the whole backlog since its saved position, as far back as your plan's history reaches, at least one request per 100 statements; delete the cursor file to start afresh instead.

for s in fp.poll_feed(sector="Energy", importance_min="high", cursor_file="energy.cursor.json"):
    print(s["indexed_at"], s["statement"], s["source_url"])

The poller pauses 150 seconds between polls unless you pass interval=. Every poll is one request (and so is every retry), so at the default a poller that runs all month makes about 17,000 requests: set interval= to fit your plan's monthly quota (a month has about 2.6 million seconds; divide by the requests you want to spend on polling). Timeouts, server errors and a passing 429 (the plan's rate limit, or too many requests at once) are retried for as long as they last, with a pause that grows to 5 minutes, and each retry is logged as a warning on the finance_pulse logger. A used-up quota, or a 429 that asks to wait more than 5 minutes, is raised at once.

More than one page

rows = list(fp.iter_statements(symbol=["NVDA", "AMD"], importance_min="medium", max_items=1000))

Each page of 100 is one request (and each retry one more). Without max_items it stops after 500 rows; max_items=None reads to the end of the data.

Errors

from finance_pulse import FinancePulseError

try:
    fp.statements(symbol="S&P")
except FinancePulseError as e:
    print(e.status, e.title, e.detail, e.param)
    # 400 Invalid symbol 'S&P' is not a ticker symbol

A 429 means too many requests: the plan's quota is used up (e.quota_used_up is true), the plan's rate limit was reached, or too many requests are open at once; the last two usually pass on their own. Send requests one after another rather than many in parallel. Timeouts and network failures are raised as FinancePulseError too, with status 0; e.transient is true for those and for 5xx answers. e.retry_after says how long to wait when the API gives a time.

Single calls raise every error at once. iter_statements retries a page up to five times in a row after a temporary failure or a passing 429, with a growing pause; poll_feed keeps retrying those for as long as they last (see above). Both raise at once a used-up quota and a 429 that asks to wait more than 5 minutes.

Methods

Parameters, response fields and limits are documented in the API reference, and the guides show complete worked examples. fp.get("/v2/...", **params) calls any endpoint the client has no method for yet.

What the data is, and is not

• Symbols are canonical entity codes with a kind. Equities and ETFs appear under their ticker (NVDA, 0700.HK); indices, rates, currencies, crypto, commodities and organisations under short codes (SPX, USD, BTC, FED). Use kind=["equity", "etf"] for tradable tickers only, and symbols(q="name", period="window") to find a code.
• Sentiment describes the statement, not a price forecast. A symbol's sentiment is the count of its statements' labels.
• Topics and storylines are grouped automatically and can be wrong or incomplete: a reading aid, not a count of events.
• How far back you can read depends on your plan. window_start in every response says where your history starts and meta() gives it in days (window_days); older statements are not returned.

See the docs for the full list of caveats.

Development

pip install -e ".[dev]"
pytest

The tests use canned responses and make no API requests.

Releasing

1. Set the new version in src/finance_pulse/__init__.py and in server.json (two places).
2. Build and upload to PyPI (uv build && uv publish). From GitHub Actions, use pypa/gh-action-pypi-publish v1.14.2 or newer: older versions reject the metadata version this build writes.
3. Only then run mcp-publisher publish: the registry verifies the package on PyPI (it looks for the mcp-name line in its description), so the release must be there first.

Licence

MIT