Next free-source huntWe check for new free data every weekday.

Data Agreement Index

Is free data correct?

Uptime monitors check whether an API answers.This checks whether the number is right: values that free MCP servers returned, next to the official publisher's own figure for the same date.
75 values compared6 of 6 datapoints measuredOfficial values fetched 2026-11-17Probe 2026-11-17: 168 servers

Datapoints

Servers measured so far

Every other audited server is not yet measured. That says nothing about whether its data is right.

How it works

Values. We only use figures a server actually returned. Most come from the weekly agreement probe: for each free, audited server whose tools look like they could serve a datapoint, it opens a normal MCP session and calls one tool that says it is read-only, once per datapoint, with fixed inputs (EUR to USD, GBP and JPY; series ids CPIAUCNS or CUUR0000SA0, EFFR and DGS10 where a tool takes FRED or BLS ids). A few come from recorded agent tests. The figure is read from the reply only when it sits in exactly one place, and is published with the inputs and an excerpt of the reply. Nothing is estimated or filled in.

Manners. The probe sends no credentials, skips servers that ask for a key or sign-in, waits at most 15 seconds per request and retries once, talks to one server per host at a time, waits two seconds between calls to one server, and makes at most one call per datapoint. Tools that do not say they are read-only are never called, and a tool whose required inputs cannot be filled with certainty is not called at all.

Official reference. The publisher's own free API, fetched by a script and committed with the date: the ECB Data Portal for euro reference rates, the BLS public API for CPI-U, the New York Fed Markets API for EFFR, and Treasury's daily par yield curve file.

Same date. We compare with the official figure for the date the server states. When a server gives only a time, we use the fixing in effect at that time (for ECB rates, the latest one published by 16:00 Frankfurt time).

Rounding. A value agrees when the gap is within what rounding can explain: half a unit in the last decimal of each figure, carried through when a server quotes the other way round (for example euros per dollar). Within 0.1% is “close”; more is “disagrees”. More than 20% away usually means a different unit, pair or series (for example a rate against the lira), so it is shown as a mismatch for a person to check and left out of the score. Servers that quote live market rates rarely match a once-a-day fixing exactly; that is a difference in what they publish, which the gap shows.

Score. Values that agree divided by values compared, per datapoint and per server, with the count shown. One value is a sample, not a track record.

Coverage. The weekly audit lists tools but never calls them. Values come from a separate, opt-in weekly agreement probe: for each candidate server and datapoint it calls one tool that says it is read-only, with fixed inputs (for example EUR to USD), and records the figure, its stated date and a short excerpt. A candidate stays not yet measured when it has no matching read-only tool, needs inputs the probe cannot fill with certainty, asks for a key, fails, or returns a reply where the figure is not in exactly one place.

Left out on purpose.
  • Stock index closes (S&P 500, Dow, Nasdaq): The index owners do not offer a free official API that allows reuse, so there is no official reference we can fetch legally. ETF prices such as SPY are not the index.
  • Crypto prices: There is no official publisher; every venue prints its own price.

Agreement compares values free MCP servers returned to Signal Lab's weekly agreement probe and recorded agent tests with the official publisher's free API, after allowing for rounding. Small samples: one recorded value says nothing about every day. Not yet measured means no value has been recorded, not that the server is wrong. Agents: /data-agreement.json or the MCP tool get_data_agreement.