regintel-mcp
Regulatory intelligence from FDA regulations.gov dockets: who commented, what they asked for, and what changed between the draft and final guidance.
Black Swan Causal Labs · blackswancausallabs.com
Retrieval lives in the MCP server. Judgment lives in skills. Humans sit at the checkpoints.
| Layer | Component | Status |
|---|---|---|
| Data | regintel-mcp server (this repo, MIT) |
Phase 1 built. Fetch, cache, extract, diff. No interpretation. |
| Method | skills: submitter-taxonomy, comment-theming, draft-final-concordance |
Built |
| Orchestration | /docket-analysis <docketId> with 4 human checkpoints |
Built; first run on FDA-2020-D-2307 |
| Output | Docket Lens explorer (reads one analysis.json; PolyForm Noncommercial) |
Built for FDA-2020-D-2307 |
Install
Python 3.11 or later.
git clone https://github.com/Black-Swan-Causal-Labs/regintel-mcp.git
cd regintel-mcp
python3 -m venv .venv && .venv/bin/pip install -e .
cp .env.example .env # then paste your key into .env
API key: get a free key at https://open.gsa.gov/api/regulationsgov/ and put it in .env as REGULATIONS_GOV_API_KEY=.... .env is gitignored. The key is sent only as the X-Api-Key header and never appears in logs, the cache or any output. DEMO_KEY is fine for single lookups but too rate-limited for a full corpus pull. Personal keys allow 1,000 requests/hour.
Register with an MCP client. MCP is an open standard, so the server works with any compatible assistant (Claude, OpenAI Codex, others). For Claude Code, from the repo directory:
claude mcp add regintel-mcp -- "$(pwd)/.venv/bin/python" -m regintel_mcp
When installed as a Claude plugin, .mcp.json starts the server from the plugin root.
Tools
Every tool reads the local SQLite cache (data/cache.sqlite) first. It only touches the network on a cache miss or when called with refresh=true.
| Tool | Does |
|---|---|
search_dockets(query, agency="FDA", docket_type?, limit=25) |
Keyword search over dockets |
get_docket(docket_id) |
Docket metadata and total comment count |
list_documents(docket_id) |
Notices, draft and final guidance, meeting material, with every file URL |
list_comments(docket_id, include_withdrawn=false) |
All comments, paginated internally. Flags is_stub, duplicate_of and agency_duplicate_count (flags only; nothing is dropped) |
get_comment(comment_id) |
Body, submitter fields, attachments and extracted attachment text |
build_corpus(docket_id, with_attachments=true) |
Pulls every comment and attachment into the cache. Resumable. Reports stubs, duplicates, OCR needs and failures |
get_guidance_text(ref) |
A guidance document (doc ID or URL) split into numbered sections |
diff_guidance(draft_ref, final_ref) |
Section-aligned draft-vs-final diff with paragraph changes and moved-paragraph detection |
export_corpus(docket_id) |
Writes output/<docket>/corpus.jsonl for skills and subagents |
For long pulls outside a chat: .venv/bin/python -m regintel_mcp build-corpus FDA-2020-D-2307.
How the diff works: the text layer is extracted with pdfplumber. Running headers, draft line numbers, footnotes and the table of contents are removed. Sections are split on the guidance's own Roman, letter and number headings, then matched draft→final by heading (parent-qualified first) and then by heading plus text similarity. Paragraphs are diffed within each matched section, and paragraphs that moved to a different section are labelled. Everything is deterministic: there are no model calls in the server.
API behaviour verified against the live service is recorded in docs/api-notes.md.
Docket Lens (the explorer)
Docket Lens (explorer/template.html) is a single page that reads one analysis.json and shows three linked views: who spoke, what they asked for (a themes × submitter-type heatmap that opens the verbatim requests), and what changed (draft vs final by section, with each request's call). Build a docket's copy with:
.venv/bin/python scripts/build_analysis.py FDA-2020-D-2307
.venv/bin/python scripts/build_explorer.py FDA-2020-D-2307
Then open explorer/FDA-2020-D-2307/index.html through any static server, e.g. python3 -m http.server --directory explorer/FDA-2020-D-2307. The explorer is licensed separately under PolyForm Noncommercial 1.0.0 (explorer/LICENSE, explorer/NOTICE).
Hosting a Docket Lens demo (Cloudflare Pages, password-protected)
sh scripts/build_site.sh FDA-2020-D-2307
cd site && npx wrangler pages deploy public --project-name docket-lens
site/functions/_middleware.js puts HTTP Basic Auth in front of every request, including analysis.json. Credentials are Pages secrets, never files: npx wrangler pages secret put DEMO_USER --project-name docket-lens, and the same for DEMO_PASS. If either secret is missing, the site serves nothing (it fails closed). It is a demo gate, not access control for sensitive data.
Status (3 Oct 2026)
- First full run: FDA-2020-D-2307 (60 comments, 1,211 requests, 4 checkpoints, human spot-check agreement 91% on 55 calls). Outputs in
output/FDA-2020-D-2307/, decisions indecisions/. - Second-docket smoke test: FDA-2021-D-0548. Retrieval and the draft/final comparison passed without changes; see
docs/smoke-test-FDA-2021-D-0548.md.
The concordance caveat
This tool measures concordance, not impact. Changes between draft and final guidance also come from FDA's internal review, ICH M14, other guidances and the evolving science. When the final text moves in the direction a commenter asked for, that does not show the comment caused the change. Every output from this project carries that caveat.
License
Server: MIT (see LICENSE). The Explorer template is licensed separately under PolyForm Noncommercial 1.0.0.
© 2026 Black Swan Causal Labs · https://blackswancausallabs.com
Official MCP Registry packaging
The Docket Lens retrieval server has an MCP bundle for Registry distribution.
See Registry installation and release instructions.
The bundle requires uv, your Regulations.gov API key, and a writable working folder.