The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Acquisition Forecasts listing page.
US federal agency procurement forecasts — what an agency has published its intention to buy, months to years before any solicitation exists. Keyless.
Part of Pipeworx — an MCP gateway connecting AI agents to 1683+ live data sources.
This is the forward-looking surface. usaspending is money already obligated;
govcon-intel scans SAM for solicitations already open. A forecast is the only
one that answers "should I be staffing for this" rather than "did I miss it".
forecast_search(keyword?, agency?, sub_agency?, naics?, set_aside?, fiscal_year?, min_value?, limit?)forecast_detail(forecast_id, agency?) — full narrative, PoP, contactsforecast_agencies() — what is wired, and the verdict for everything that isn'tAgencies are required to publish forecasts and acquisition.gov indexes 22 of them. I probed every one on 2026-08-29. Exactly one publishes machine- readable data:
| Verdict | Count | Agencies |
|---|---|---|
| JSON API | 1 | DHS (apfs-cloud.dhs.gov, 847 rows) |
| HTML, no data file | 15 | Agriculture, Education, Energy, Justice, Labor, State, Interior, Treasury, Transportation, VA, GSA, NASA, NSF, SSA, DHS's own landing page |
| PDF only | 4 | DoD, HUD, NRC, OPM |
| Blocks automated clients (403) | 2 | USAID, Commerce |
| Unreachable (timeout) | 1 | HHS |
PDF forecasts are deliberately not OCR-ed. A parsed number that is wrong is worse than no number, and these are dollar values people plan against.
Because of that, forecast_agencies() ships as a first-class tool and
agencies_covered + coverage_note ride on every search response. A pack named
for federal forecasts that quietly covered one agency would answer "nothing
planned" for 21 agencies it never checked — the absence-of-data-vs-absence-of-
the-thing failure, which is the one this catalog keeps re-learning.
Adding an agency is a config entry in AGENCIES[] plus a field map.
The APFS API ignores every query parameter. ?naics=5415,
?organization=CBP and ?zzznonsense=x all return the identical 847 rows with
the identical first id. There is no server-side filtering, so the pack fetches
the whole file (~2 MB, cached 15 min) and filters in memory. Anyone
"optimising" this by pushing filters upstream gets a result that looks filtered
and is not.
small_business_set_aside and small_business_program are different fields.
The first is a boolean (true/false/null), the second is the program name
(SB, 8(a), HUBZone, WOSB, SDVOSB, AbilityOne, None, TBD).
Coalescing the boolean first turns 350 of 847 rows into the literal string
"true" and costs them their program name — and makes set_aside="8(a)" return
27 rows instead of 71, because only rows whose boolean happened to be null
could still match. They are returned as set_aside (the name) and
set_aside_applies (the boolean, null meaning the row does not say).
NAICS arrives as code and label in one string — "236220 - Commercial and Institutional Building Construction". Matching a caller's 5415 against the raw
field would also hit any description containing those digits. The code is split
out and matched as a prefix, which is how NAICS hierarchy works: 5415
covers 541511, 541512, … (175 rows vs 19 for the full code).
Values are bands, not numbers — "$50M to $100M", "Over $100M". The band
text is preserved verbatim and parsed to low/high. For an open-ended top
band high is null, not a large number, and min_value treats it as
unbounded — a fabricated ceiling would be sortable and wrong.
organization is a sub-agency path, not an agency: CBP/OIT/LSS/ITB. All
847 rows are DHS. agency and sub_agency are reported separately so
agency="DHS" means what a caller expects.
Forecasts are revised. published_date, previously_published and
last_updated say which version you are looking at; a withdrawn or awarded
requirement can drop out of the file entirely, which is why forecast_detail
says so rather than just reporting "not found".
Add to your MCP client (Claude Desktop, Cursor, Windsurf, etc.):
tools/list at https://gateway.pipeworx.io/acquisition-forecasts/mcp returns the tools in the table
above plus the shared Pipeworx meta-tools — ask_pipeworx,
discover_tools, search_within, remember/recall and the rest of the
gateway-wide set. So the tool count you see is larger than this table: a
single-pack endpoint currently lists roughly 30 shared tools alongside the
pack's own. The connection's initialize response states its exact scope, and
is the authoritative answer for a given day.
This is deliberate, not multiplexing by accident. The meta-tools are what let a
scoped connection answer a question this pack does not cover — via
ask_pipeworx, which routes across the whole catalog — without you adding a
second MCP server. There is currently no way to mount a pack endpoint without
them; if the extra schemas cost you more context than the routing is worth,
connect to the full gateway once rather than to several pack endpoints.
Or connect to the full Pipeworx gateway to get every pack's tools listed directly, instead of just this one's:
Both URLs reach the same gateway and the same 1683+ data sources. The
only difference is which pack's tools are listed directly; ask_pipeworx
reaches all of them from either one.
No account needed for the first calls. Inspect any tool: GET https://gateway.pipeworx.io/v1/tools/forecast_search. Find one: POST https://gateway.pipeworx.io/v1/tools/search_packs with {"query":"..."}.
This package also runs as a local stdio MCP server — no Pipeworx account, no gateway round-trip:
Or run it directly to confirm it starts:
It speaks MCP over stdin/stdout and answers initialize/tools/list/tools/call
for only this pack's tools — none of the shared meta-tools the gateway
connection above adds. Same source, same tools, no ask_pipeworx routing.
Instead of calling tools directly, you can ask questions in plain English — this works on the pack endpoint above as well as on the full gateway:
The gateway picks the right tool and fills the arguments automatically.
MIT