ChatGPT Importer

August 24, 2026 · View on GitHub

Import a ChatGPT conversation into local files: capture it from the browser, normalize it, and render Markdown, JSON or a styled PDF with user and assistant turns marked, artifacts identified and conversation metadata recorded.

Built for the case where a ChatGPT thread holds decisions or background that a repository does not, and the point is to get it somewhere the next agent will read it.

Pipeline

capture ──> raw JSON ──> normalize_conversation.py ──> normalized JSON ──┬─> render_conversation.py  (Markdown)
                                                                        ├─> render_pdf.py           (PDF via Typst)
                                                                        └─> redact_conversation.py  (then render)

The normalized JSON is the artifact worth keeping. It holds turn roles, timestamps, artifacts, citations and stats together, and every renderer reads it.

Capture routes

Input	Route	Completeness
`chatgpt.com/c/<id>`	Authenticated backend API, run in the page	Complete
`chatgpt.com/share/<id>`	Same, no token	Complete
Export `.zip`	Local file, no browser	Complete, plus real image files
API route broken	DOM scrape with scrolling	Lossy — no timestamps, branches or artifact payloads

Skills

Skill	Does
`import-chatgpt-conversation`	Entry point — picks the route, runs the pipeline
`conversation-to-context`	Import into `context/` in the current repo as agent context
`capture-via-api`	Primary browser capture
`capture-via-dom`	Fallback scrape
`import-export-archive`	Bulk import from an account export
`render-markdown`	Markdown with front matter, stats and sources
`render-typst-pdf`	Styled PDF, configurable speaker identity
`redact-conversation`	Two-pass redaction
`extract-sources`	De-duplicated citation list
`extract-artifacts`	Save canvas documents and code as files

Metadata recorded

Every import captures the date the conversation was held, its duration, model, total turns, user and assistant message counts, character counts per speaker, artifact and citation counts, and whether the thread contained abandoned edit branches.

Requirements

Capture: Chrome with the Claude in Chrome extension, logged in to ChatGPT.
Normalize and Markdown: Python 3.9+, stdlib only. No venv.
PDF: pandoc >= 3.0 (for the typst writer) and typst.

Privacy

Conversations are private by default and nothing captured is committed. The .gitignore covers the usual output paths (imports/, out/, context/ transcripts, conversations.json).

Captured bodies are written to disk rather than returned through the tool result, so a long private thread does not end up copied into an agent transcript. Only a summary comes back.

redact-conversation runs a validated pattern pass — keys, emails, IBANs, Luhn-checked cards, check-digit-verified Israeli IDs, phones, IPs — and then prompts for the judgement pass that regex cannot do.

Format notes

The ChatGPT conversation format is reverse-engineered and undocumented. references/chatgpt-conversation-format.md records what was verified against a live account on 2026-08-24, what was inferred, and the traps — chiefly that mapping is a DAG containing abandoned branches, and that the rendered DOM is virtualised so a single-pass scrape silently returns a fraction of the thread.

Testing

./tests/run.sh

Runs the whole pipeline against fixtures/sample-conversation.json, a wholly synthetic conversation exercising branching, hidden messages, canvas artifacts, image pointers, citations and an unknown content type.