Check the regulation in the conversation you’re already in.

Your assistant answers, and then you go find the actual language, because you’re the one signing off. This server ends that round trip: the section itself, verbatim, with its citation, the date it took effect, and the document it came from.

What you need: Claude on any plan (the free plan allows one custom connector), or ChatGPT Plus, Pro, Business, Enterprise or Edu. No account with us, no API key, nothing to download.

Why this exists

We measured what AI assistants cite for FHA rules. Across every citation of Handbook 4000.1 where the edition could be identified, 166 of 166 pointed at a superseded version. The most-cited edition was from October 2023; the newest anything cited was May 2024. Most came from HUD’s own website, which still serves old PDFs that look current.

The models weren’t misreading the text. They were reading the wrong year, and nothing in the answer said so.

Which assistant works best

We tested this server with real loan-file questions, asked the way loan officers and processors actually ask them. The model you use matters more than anything else:

Assistant	How it does
Claude any current model	Recommended. Finds the governing section even from shorthand (“DTI”, “comps”), quotes it verbatim with the citation, and usually checks its own quotes against the regulator’s document before answering.
ChatGPT current GPT-5 models	Recommended. Just as accurate and very thorough: it verifies every quote, so answers take longer, often a minute or more.
Older or lighter models GPT-4o class, “mini” models	Use with care. Usually finds the right section, but tends to summarize instead of quoting, and a summary can’t be checked. Ask it to quote the text and cite the section (tips below).

One rule for every assistant: an answer without a verbatim quote and a section citation is unverified. The server gives the assistant everything it needs to quote; hold it to that.

What your assistant gets

search_regs: Finds the governing section by topic, in plain words or the regulator’s own. Understands loan-file shorthand. Weak matches are labeled as partial, and questions outside mortgage regulation usually get no results rather than a stretch.
lookup_citation: Goes straight to a citation you already have: B3-4.3-04, 4000.1 II.A.4.c, 12 CFR 1026.43. When a citation is genuinely ambiguous, it says so instead of guessing.
get_reg · get_obligation: The verbatim text, the section path, the effective date, and the dated snapshot it came from.
verify_quote: Checks whether a passage appears in the regulator’s document: exact, normalized, assembled from separate passages (with anything skipped shown), or not found. Works on text from anywhere, including your own draft.

What you can run

Beyond lookup, the server carries four playbooks: guided compliance work that runs in your chat, grounded in the same verbatim text. Start one by name once the server is connected.

FNMA AI Lender Letter: LL-2026-04: find the AI in your P&Ps, then draft the AI system inventory, risk assessment, governance policy, P&P updates and change log. Say: Run the FNMA AI Lender Letter playbook on my P&Ps.
FHA underwriting QC: Pre-fund QC of an FHA loan file against Handbook 4000.1: underwriting decision, credit history, property and valuation. Say: Run the FHA underwriting QC playbook on my loan file.
VA loan QC: QC of a VA loan file against the Lenders Handbook (M26-7): eligibility and entitlement, residual income and DTI, property, and IRRRL or cash-out refinance rules. Say: Run the VA loan QC playbook on my loan file.
USDA loan QC: QC of a USDA guaranteed loan file against HB-1-3555: household income limits, credit, repayment ability, property and site. Say: Run the USDA loan QC playbook on my loan file.

All playbooks →

The lookup tools are not interpretation: they hand your assistant the rule and where it came from. The playbooks produce working drafts for your review, not legal advice.

Install

Server address

Claude desktop app or claude.ai

1
Open Settings, then Connectors in the sidebar.
2
Click Add custom connector.
3
Name it Claude for Compliance and paste the server address above.
4
Under Authentication, choose No sign-in (Claude detects it automatically). Leave request headers empty, then click Add.
5
In a new chat, switch the connector on from the tools menu and ask a regulatory question, or pick one of its playbooks from the prompts list.

Works on every plan; the free plan allows one custom connector. On Team and Enterprise, an owner may need to allow custom connectors first. Claude Desktop won’t pick up a remote server from claude_desktop_config.json: add it through Connectors.

ChatGPT Plus, Pro, Business, Enterprise, Edu

1
Open Settings → Apps & Connectors → Advanced settings and turn on Developer mode.
2
Back in Apps & Connectors, click Create.
3
Name it Claude for Compliance, paste the server address above, and set authentication to none.
4
In a new chat, make sure the connector is switched on, then ask a regulatory question.

Not available on the free ChatGPT plan. On Business, Enterprise and Edu, a workspace admin may need to enable Developer mode first. If the tool list ever looks stale, open the connector and click Refresh.

Claude Code

claude mcp add --transport http claude-for-compliance https://mcp.claudeforcompliance.com/mcp

Cursor and VS Code one click

Add to Cursor Install in VS Code

Each opens an approval dialog with the server filled in.

Once it’s connected, a good first question:

Using the Claude for Compliance tools, quote the verbatim text of Fannie Mae Selling Guide B3-4.3-04 (Personal Gifts) with its section citation and effective date.

Getting good answers

Ask for the text, not just the answer. “Quote the exact language and cite the section” turns a summary into something you can check.
Name the agency. “Fannie”, “FHA”, “VA”: the same topic often has different rules at each.
Watch for special programs. A limit from a program like HomeStyle, a high-LTV refinance, 203(k) or an IRRRL applies only to that program. If the answer cites one, ask whether it’s the general rule.
Check the date. Most sections carry an effective date. If the answer doesn’t mention one, ask for it; if the section has none, the server says so.
Have it check itself. “Verify those quotes” runs each one against the regulator’s document.

No connector? The corpus is still yours.

Free ChatGPT can’t add connectors, and some workplaces don’t allow them. Everything the server returns is also published free: the regulation pages themselves, per-section CSV and XLSX files, and full-library bundles you can attach to any assistant.

Browse the corpus · Full library (CSV, zipped)

What the server sees

The server receives only what your assistant sends its tools: searches, citations, and passages to check. It never sees the rest of your conversation. There are no accounts, and we don’t store anything that identifies you. Searches that find nothing are kept as the search text alone, so we know what to add to the corpus. Everything the server returns is public regulatory text.

Limits and support

Read-only. Every tool looks things up or checks text. Nothing the server does changes, sends or stores anything on your side.
No sign-in. No account, no API key, no OAuth screen.
Rate limit: 10 requests per 10 seconds from one network address. A busy assistant can hit it; wait a few seconds and ask again.
Daily budget: 100,000 requests a day across all users. If it runs out, the server says so and resets at 00:00 UTC.
No IP addresses stored. Usage counts record only which kind of client connected (Claude, ChatGPT, Cursor and so on).
Not legal advice. Every result carries a notice: for research purposes only, provided as-is, verify against the primary source. The server returns the regulator’s text and where it came from; the judgment stays with you.
Support: email earl@cfsynergies.com or use the contact form. See the terms and privacy policy.