An MCP server for Claude, and for any other platform that speaks the protocol. It answers filing questions in typed lanes, and links every result back to the filing.
Four failures, each one recorded against a real filing while this corpus was being built. None of them announced itself.
One SpaceX free writing prospectus arrived as 443 documents: 442 JPGs and a single HTML file. The passage that mattered sat on image 191 of 442, in a picture, where no text index could reach it.
Full-text search tells you the words are in the filing somewhere. It does not tell you which page, and when the disclosure lives inside a JPG, it cannot see the words at all.
Apple reported net sales of $416,161 million. iPhone revenue for the same year is exactly $209,586,000,000, and calling the SEC's company facts API does not return it. That feed serves only default-dimension facts, so the per-product breakdown is stripped out before the response reaches you and the number never comes over the wire.
A figure that passes through a summarizer is a figure somebody has to check.
Apple's Item 1A ran 68,767 characters one year and 68,052 the next. A plain paragraph diff over the two reported 20 additions and 20 removals, and left exactly one of 48 clause pairs unchanged.
Almost nothing changed. Almost everything moved.
A holding-company reorganization handed the registrant that trades as APO a brand new CIK in May 2021, opened as a merger shell named Tango Holdings, and it has four annual reports to its name. Ask for fifteen years of Apollo revenue and a search without a curated lineage returns six rows from that four-year company. The nine earlier years, FY2011 through FY2019, sit on a predecessor CIK that carries no ticker at all.
A wrong CIK does not raise an error. It answers.
Take the ten highest-ranked additions to Apple's risk factors between fiscal 2024 and fiscal 2025. Then check each one against the prior year's text by hand.
Seven appeared word for word a year earlier. One was reworded. Two were genuinely new.
The diff was not lying. It was answering which spans are new, and that is not the question the lawyer asked. When a filing is reflowed, last year's document is the closest possible distractor to this year's, so the corpus pairs clauses first, demotes the moves and renumbering, and leaves the model to reason over what actually survived the comparison.
The words that really moved are countable, and counting them is the check. Stock price appears 7 times in the fiscal 2024 item and 38 times in fiscal 2025. Tariff goes from 2 to 10. Ransomware appears 4 times in each year, which is why a confident summary about a strengthened cybersecurity risk factor read so well and was wrong.
A question about a number, a question about a change, and a question about language are three different retrieval problems, and one search box answers all three the same way. This corpus routes them apart before it looks for a result. Financial facts resolve against XBRL in SQL. Changes resolve against a clause-paired diff. Language resolves against two text lanes, one over filing bodies and one over press material and slide decks, fused and then reranked.
Even inside the language lane the two retrievers disagree. Ask about GPUs bought on a purchase-order basis with no long-term supply contract and pure semantic search answers with a cloud services agreement, the contract under which SpaceX is paid for supplying compute rather than the risk of buying it. It has the topic right and the direction of the money wrong. Ask the same thing the way a lawyer would, about reliance on outside parties for hardware, and pure keyword search answers with the company's plan to build its own chips, which is the closest thing in the filing to the opposite of the question.
Neither the semantic nor the keyword retriever is broken. They fail in opposite directions, which is why the balance between semantic and keyword search, the hybrid dial, is set per question by the LLM during the tool call, and not once by whoever configured the server.
Unknown is not zero. When an issuer never tagged a metric, the answer is that it is not reported, not that it is nil.
Each one answers a question the others answer badly.
Deterministic XBRL, queried in SQL over more than six million financial facts, including the dimensional facts by product and by segment that the consolidated feed leaves out. No figure on this path passes through a summarizer or a similarity ranking.
Clause-paired redlines, deletions struck through and insertions underlined, in the form attorneys have read redlines in their whole career. Reflow, renumbering and cosmetic edits are demoted before you read the result. The method runs in three steps and the tools hold the order: a summary contrast finds what moved, the redline supplies the clause you can quote, and counting the distinguishing language on both sides decides whether it changed or was only reorganized.
Dense and keyword retrieval fused by rank and then reranked, across filing bodies and press material together. The balance between them is a dial, because each retriever fails precisely where the other one works, and the dial is turned per call by the LLM, which is the only party that knows whether it just typed a defined term or a theme.
Section 16 insider transactions, 13F institutional holdings, Schedule 13D and 13G stakes, and proxy proposals, held as tables you can filter, sort and join. Roughly 955,000 insider transactions and more than 18 million institutional holdings sit alongside the financial facts.
Image-only exhibits and investor slide decks are read with OCR and a vision-language model and indexed page by page, so an answer can name slide 191 of 442 and hand you the link to it. A disclosure that exists only as a chart is still a disclosure.
Companies resolve to a CIK before anything else runs, with predecessors, subsidiaries and acquired entities grouped by hand. Broadcom is three CIKs and one continuous business, running back through Avago, and no field in EDGAR records that the three are the same company. Somebody has to write that down and keep it current. A fifteen-year question against a company that has only filed for four years gets told so.
The assistant picks the lane, out of what each tool says about itself: which call to make before this one, and what the result means once it arrives. The tool text is what to do first. The answer carries what to do next.
You can see which lane it picked, and follow the result back to the filing.
Every result carries the filing it came from. Search results are ranked by similarity and not by authority, and they say so on the result itself rather than in a footnote nobody reads.
Retrieved filing text is treated as data, never as instructions. Exhibits can embed arbitrary third-party documents, so text that reads like a directive is flagged and reported as quoted content rather than acted on. Nothing is deliberately stripped or rewritten, because legal research depends on verbatim fidelity and a cleaning pass would quietly corrupt the covenant you are citing; where text arrives through OCR or a vision model, it is kept as close to the page as those lanes allow. The flag names the document and the kind of hit, never the matched string, and it is absent altogether on a clean result, because a caveat attached to every answer is a caveat that gets skipped.
Issuer-drafted language is presented as attributed advocacy. A prospectus describing its own supply chain is the issuer's characterization of it, not an established fact. Issuers know their disclosures are read by machines now, and phrasing a risk factor so that it survives summarization intact is drafting rather than an attack. The answer to it is attribution, not detection.
The tool declines rather than blends. Ask for free cash flow across NVIDIA and Apple in one call and it refuses and tells you to ask once per company, because pairing one issuer's operating cash flow with another's capital expenditure produces a figure belonging to neither, out of two inputs that were each parsed correctly from a real filing.
The question is treated as the sensitive part, because it is. The filings are public; what a firm is asking about is not. Tool arguments are never written to a log, a link to a full result set is encrypted rather than merely signed and stops working after five minutes, and a firm that would rather its queries never reached anyone's vendor can point search at its own embedding endpoint, bring its own reranker, or switch reranking off.
The research corpus covers the top 175 US public companies by market capitalization, with substantive filing history back to mid-2009 (the start of the SEC's structured-data era). High-volume administrative filings and structured-product pricing supplements are excluded to keep search results relevant. Any other company can be added to the corpus for a small one-time price. See the full list.
$39/month
Auto-renews at the then-current price until cancelled. Cancel any time from your account's billing portal; cancellation stops future renewals and does not itself refund the current period.
Private beta. Access is by invitation: [email protected].
EDGAR Intelligence is a research tool over public SEC records. Summaries and rankings are AI-generated and require professional review. Verify against the linked source filing before relying on any output. The explainer videos are AI-generated as well: each one was created with Google's Gemini Notebook, formerly NotebookLM, as an overview of the article it accompanies, and it carries the same review caveat. The articles themselves are written.
