Acquisition.gov MCP

Read-only, deterministic MCP access to the official Acquisition.gov FAR Overhaul (RFO) model-part pages, the posted agency-deviation index, official deviation PDFs, and a small allowlist of RFO guidance resources.

This server reports source content and metadata. It does not decide which rule governs a procurement. In particular, model deviation text is not treated as operative for an agency without that agency's posted deviation.

Install

uvx acquisition-gov-mcp==1.0.8

The server uses stdio, requires no credentials, and defaults to a three-second cross-process interval between Acquisition.gov requests. FEDERAL_API_MIN_INTERVAL_SECONDS may increase or decrease that interval for controlled testing; production clients should retain three seconds.

Request pacing

The next request starts after the previous request's duration plus 3 seconds. This is a completion delay, not a 3-second start interval. Local processes sharing the same pacing directory and identity share this gate; a separate local counter does not create additional provider quota.

See the complete pacing reference for all nine servers, shared credentials/IPs, configuration and hosting differences.

The hosted Acquisition.gov endpoint retains 60 HTTP requests per 60 seconds per incoming IP and Cloudflare location, 4 active MCP HTTP requests, a 55-second backend processing timeout, and 64 KiB request bodies. Its one-upstream-request-at-a-time gate is shared by all hosted users.

Tools

Every retrieved source includes a canonical URL, UTC retrieval time, SHA-256 content hash, extraction status, and warnings. Agency PDF dates and applicability are returned only when labeled language is found inside the document; filenames are never used to infer them.

Safety boundary

• Only https://acquisition.gov and https://www.acquisition.gov are allowed.
• Redirect targets are revalidated; credentials, explicit ports, arbitrary hosts, and private IP targets are rejected.
• Downloads are bounded to 5 MiB HTML / 25 MiB PDF. HTML complexity, redirects, page selection and output length are also bounded.
• PDF extraction runs in one isolated parser subprocess at a time, with a 30-second deadline, 2 MiB decoded page-stream limit, and 160 MiB address-space / 8-second CPU limits on Linux. macOS and Windows retain process isolation and explicit content/output limits without the Linux resource caps.
• HTTP 429 is not burst-retried. Retry-After is retained in the shared pacing state.
• If the Python TLS transport stalls against Acquisition.gov's CDN, the server may use an installed system curl for the same prevalidated URL; redirects remain disabled and revalidated by the server.
• Duplicate and conflicting index entries are returned with warnings instead of silently resolved.
• Scanned, encrypted, and malformed PDFs return explicit extraction status and metadata where possible.

Version 1.0.8 passed 236 offline tests, including 55 independent-review regressions, real stdio/HTTP smoke checks, parser isolation, cancellation and per-tool correctness cases. See testing.md for evidence, exact limits, known scope and reproducible commands.