ALFRED

Two read-only tools over ALFRED, FRED's real-time (vintage) view. Instead of today's revised figure, ALFRED reports what a series said as of a historical publication date.

A real example, verified live against FRED: CPIAUCSL for March 2020 was first published as 257.953 (vintage 2020-04-10), then revised to 257.989 (vintage 2021-02-08), and revised again since. A plain FRED pull silently substitutes today's fully revised number for "the" March 2020 value. ALFRED keeps every vintage, each tagged with the date it was known on.

from lazybridge import Agent, LLMEngine
from lazytools.connectors.alfred import ALFREDTools

agent = Agent(name="macro", engine=LLMEngine("deepseek-v4-flash"),
              tools=[ALFREDTools()])

The two tools

Tool	Answers
alfred_vintage(series_id, as_of, start="", end="")	What was this series publishing on as_of?
alfred_vintage_dates(series_id, limit=200)	Which vintages exist at all, newest first?

as_of is required on alfred_vintage. Defaulting it to today would return revised data through a tool whose whole purpose is not to — the call is refused instead, before any request goes out.

Live, not warehoused

This connector calls FRED at request time and stores nothing. That is a deliberate change from its first shape, which read a table market-data-hub ingested.

Vintage data is asked for one series and one date at a time, when a backtest reaches a decision point. It is not a series anyone sweeps daily. Ingesting it would mean choosing in advance which series and which vintages might someday be wanted — and being wrong about that is invisible until the question is asked and the answer is missing. Asking at call time has no such guess in it.

The trade is the usual one for a live connector: it needs the network and a key, and it is not available to a job running offline.

Credential

Needs a free FRED API key in FRED_API_KEY — the same variable market-data-hub already resolves, deliberately not a second name for one credential. Keys are free from fredaccount.stlouisfed.org.

Resolution is lazy: importing the module and constructing the provider never require the key, so the MCP server mounts the provider either way. Only a call fails, and it fails with an instruction rather than a KeyError.

The key is never written into an error message. FRED reports a bad series id, an impossible date and an invalid key alike as HTTP 400 with an explanatory body; that body is passed through because it is the useful part, but the query string — which carries the key — is not. A test pins this.

Two failure modes worth knowing

A vintage older than the archive. ALFRED's vintage history for a series starts later than the series itself. CPIAUCSL has observations from 1947 but only 668 vintages, the oldest being 1972-07-21. Asking for a vintage before that returns HTTP 400, and FRED's own advice in the body is to remove realtime_start — which would turn a point-in-time read into a revised-data read, precisely the mistake this connector exists to prevent. That advice is therefore replaced with a pointer to alfred_vintage_dates.

A truncated answer that looks complete. alfred_vintage_dates returns the vendor's own total alongside the dates, and flags the reply as truncated when the two differ. A partial list that looks whole is how a caller concludes a series has a short history. Observations truncate at 400, keeping the newest — a vintage read is almost always asked backwards from a decision date, so the recent end carries the answer.

Why this matters for walk-forward backtests

A backtest asking "what was CPI for March 2020?" through a plain FRED pull gets whatever FRED serves today, even for a simulated decision date in April 2020 when only 257.953 had ever been published. That is look-ahead bias: the strategy sees a number nobody could have known. Pinning as_of to the decision date reproduces exactly what was public that day, which is the only version of the series a walk-forward simulation is entitled to use.