| name | capture-via-api |
| description | Capture a ChatGPT conversation as raw JSON by calling the authenticated backend API from inside the browser tab. Use when importing from a chatgpt.com/c/ URL or a /share/ link. This is the primary, complete capture route. |
Capture via the backend API
Runs a fetch in the page's own context, so it inherits the logged-in session.
Returns the complete conversation: every message, timestamps, model slugs,
tool calls, artifact payloads and all branches.
Steps
- Load the browser tools in one call:
ToolSearch "select:mcp__claude-in-chrome__tabs_context_mcp,mcp__claude-in-chrome__tabs_create_mcp,mcp__claude-in-chrome__navigate,mcp__claude-in-chrome__javascript_tool"
tabs_context_mcp for tab ids, then tabs_create_mcp for a fresh tab
unless the user asked to use one they already have open.navigate to the conversation URL. Wait for it to settle.- Read
${CLAUDE_PLUGIN_ROOT}/scripts/browser/capture-via-api.js and pass its
contents to javascript_tool.
- The snippet returns a JSON summary and writes the conversation to the
browser's download directory.
Reading the result
Success looks like:
{"ok":true,"filename":"chatgpt-conversation-<id>.json","bytes":126173,
"nodes":35,"roles":{"user":17,"assistant":17},"branch_points":0,"leaves":1}
Only the summary comes back. The conversation body goes to disk on purpose —
returning it would copy a private thread into the transcript and spend a large
amount of context on something no one needs to read in-band.
branch_points > 0 means the thread has edited or regenerated turns. The
normalizer renders the branch the UI last displayed and reports the rest in
stats. Say so when reporting, because the user may want a different branch.
Finding the downloaded file
Do not assume ~/Downloads. Chrome's download directory is a per-profile
setting and is not always the default — ~/Desktop is a common alternative.
Locate the file by name instead:
find ~ -maxdepth 3 -name 'chatgpt-conversation-*.json' -mmin -3 2>/dev/null
Move it out of the download directory before working on it — those folders get
cleaned, and the file is private.
Failure modes
| Returned | Meaning | Do next |
|---|
{} | The snippet lost its await prefix | Re-run with the file's exact contents |
not_authenticated | No accessToken from /api/auth/session | Ask the user to log in to chatgpt.com in this Chrome profile; do not attempt to log in yourself |
http_401 / http_403 | Token expired, or the conversation belongs to another account | Ask the user to reload the tab |
http_404 | Wrong id, or a deleted conversation | Re-check the URL |
not_on_a_conversation_page | Tab is on the chat list, not a thread | Navigate to the /c/<id> URL |
no_mapping_in_payload | Response shape changed | Fall back to capture-via-dom and flag that the API route needs updating |
Never enter credentials. If the account is logged out, that is the user's job.
