Reading live SEC filings from ChatGPT or Codex
Not on their own. Neither model holds filings published after its training cutoff, and Codex CLI runs sandboxed with network access disabled by default. Both can read live filings through a tool: a browsing tool, EDGAR's own endpoints with a declared user agent, or an MCP server returning filings as JSON.
/mcp. Nothing here needs an API key or a signup.
The distinction matters because the failure is quiet. Ask a model about an 8-K filed this week and it will often produce something shaped exactly like an answer, assembled from filings it saw during training and general knowledge of how 8-Ks read.
Three ways to get filings into the session
| Method | What you get | Friction |
|---|---|---|
| Browsing tool | Rendered EDGAR pages | Parsing, rate limits, sandbox off by default |
| EDGAR endpoints directly | Authoritative source data | CIK lookup, form mapping, declared user agent |
| Data wire over MCP | JSON with timestamps and links | One call, priced per call |
Browsing is the path most people try first and abandon second. Codex CLI in particular sandboxes the session and disables network access unless you enable it, so a browsing attempt fails in a way that looks like the model refusing rather than the environment blocking.
What EDGAR requires of automated callers
Going to the source directly is the right instinct and the SEC supports it, with two conditions that catch people out.
The SEC's webmaster FAQ states that the current maximum access rate is 10 requests per second, monitored to preserve equitable access, and asks callers to declare a user agent in request headers. That second condition is enforced, not advisory. Fetching sec.gov from this machine with an ordinary browser user agent returned HTTP 403; the same URL with a declared automated-tool user agent in the documented format returned 200.
Past the headers, the work is the part nobody budgets for: mapping a company name to a CIK, mapping the question to a form type, walking the rolling index files, and deciding which of several documents in a submission is the one you wanted. It is all doable and all yours to maintain.
Calling the EDGAR wire instead
The wire does the CIK and form mapping and returns ranked filings as JSON, so the agent makes one call and reads a structured answer.
# free: confirm the filing is in the window (top 3, no charge) curl "https://thebotwire.com/edgar/preview?q=Nvidia&form=8-k" # paid $0.01: ranked filings with the form filter and a rolling window curl "https://thebotwire.com/edgar/filings?q=Nvidia&form=8-k&since=7d&limit=10"
The form filter accepts 8-k, 10-q, 10-k, form-4, s-1, 13f, 6-k and 13d. since accepts windows from 30m to 14d, and q is ranked keyword search, so send distinctive terms rather than a sentence. For Codex, the same wire arrives as a tool with codex mcp add botwire -- npx -y botwire-mcp; in Claude Desktop it is a JSON entry, covered in connecting an MCP server to Claude Desktop.
What this does not give you
Worth stating plainly, because an agent that pays and gets the wrong shape of data does not come back.
- Not full filing text. You get the filing's metadata, a summary and a link to
the document on sec.gov. Reading the whole 10-K is a separate fetch.
- No XBRL extraction. Financial statement line items are not parsed out.
- A rolling window, not an archive. Sparse sources keep up to 30 days. EDGAR
itself goes back to 1994 and remains the place to go for history.
- Insider trades via Form 4, not a positions database. The filter surfaces the
filings; interpreting them is the agent's job.
If what you need is "did anything get filed about this company in the last week", the wire answers it in one call. If you need to parse a full annual report, retrieve the document from the link and handle it yourself. Per-call cost including the tokens spent reading the response is broken down in what a live data query costs an agent.
FAQ
Does ChatGPT have live SEC data built in?
No. ChatGPT can reach the web when browsing is enabled, which is not the same as having a filings feed. It retrieves and parses pages at question time, so results depend on what the search surfaced and whether the page rendered, rather than on the filing index itself.
Can Codex reach the network at all?
Yes, once you allow it. Codex CLI sandboxes the session with network access off by default, so tools that need the internet have to be permitted explicitly. An MCP server registered with Codex is the cleaner path, because the server makes the network call and returns structured data to the model.
Which form type should an agent filter on?
8-k for material events as they happen, 10-q and 10-k for periodic results, form-4 for insider transactions, s-1 for registration statements, and 13f or 13d for holdings and stakes. Filtering matters: an unfiltered query mixes routine periodic filings with the event you were watching for.
What if the wire returns no filings?
Widen since before concluding nothing happened. EDGAR is quiet outside filing seasons, and a count:0 with an old freshest timestamp on a sparse wire usually means the window was too narrow. Freshness is checkable at any time on /health.
Sources
The 10 requests per second access rate and the declared user agent requirement are stated in the SEC's EDGAR webmaster FAQ, which also notes that EDGAR's electronic filing archive begins in 1994.
Related: How much does it cost for an AI agent to query live data? · How AI agents get real-time data past the training cutoff · How do AI agents track new US federal regulations?