Convolvger

Local-first, provider-independent archival for public AI conversations.

Convolvger takes a public AI conversation share URL, reconstructs the conversation into a canonical internal model, validates the extracted data, and exports it into portable formats such as Markdown and JSON.

Install

Convolvger needs Python 3.12 or newer. The simplest route is uv, which downloads a suitable Python itself if you do not have one.

macOS, Linux, WSL

curl -LsSf https://astral.sh/uv/install.sh | sh
uv tool install convolvger

Windows (PowerShell)

powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"
uv tool install convolvger

Without uv

pipx install convolvger

or pip install convolvger inside a virtual environment running Python 3.12 or newer. Installing with an older interpreter fails with Requires-Python >=3.12, which is the error to expect on a system Python 3.10 or 3.11.

Check the install:

convolvger --help
convolvger --version

--version reports the installed version, the interpreter it runs on, the archive schema it writes and the ones it can read, and the providers this build supports. It is the quickest thing to include in a bug report, because the version it prints is the one your archives record.

Upgrade

Convolvger never updates itself and never checks for a new release. A tool that keeps your conversations on your machine has no business contacting an index to ask about itself, so upgrading is something you do deliberately, with whichever tool installed it:

uv tool upgrade convolvger      # if you installed with uv
pipx upgrade convolvger         # if you installed with pipx
pip install -U convolvger       # inside a virtual environment

Then run convolvger --version to confirm which copy you now have. Releases are listed at github.com/bilalmughal1/convolvger/releases.

Archives written by an older version stay readable: each one records the schema it was written against, and a newer Convolvger reads the versions it lists under --version rather than assuming.

Usage

Every command begins with convolvger, followed by a subcommand. There are five: providers, bookmarklet, inspect, export and verify. Note that a bare export <url> runs your shell's own export builtin rather than this tool.

See which providers are supported

convolvger providers

Look at a conversation without saving anything

convolvger inspect https://chatgpt.com/share/SHARE_ID

Prints the title, provider, message counts and every extraction finding, then exits without writing a file.

Export it

convolvger export https://chatgpt.com/share/SHARE_ID

Writes Markdown to a filename derived from the conversation title, in the current directory. To choose the format or the destination:

convolvger export https://chatgpt.com/share/SHARE_ID -f json
convolvger export https://chatgpt.com/share/SHARE_ID -o mychat.md
convolvger export https://chatgpt.com/share/SHARE_ID -o -

-f json writes the archival record; -f md (the default) writes the reader-facing document. -o - writes to stdout instead of a file.

Parse a snapshot you already saved

convolvger export https://chatgpt.com/share/SHARE_ID --from-file saved-page.html

--from-file parses a saved copy of the share page instead of fetching the URL. The URL is still required: it routes the snapshot to the right provider, and it is the provenance the archive records. Because when a saved file was captured is not knowable from the file, an archive made this way records no retrieval time rather than claiming a false one.

This is how to re-read a capture without asking the provider for it again, which matters because the answer may have changed since. It works with inspect too.

Archive a Gemini conversation

convolvger export https://gemini.google.com/share/SHARE_ID

Gemini share pages are fetched directly, with no browser involved. The page itself carries no conversation — it is an application shell that loads one afterwards — so Convolvger asks for the same data the page does. All three link forms work:

https://gemini.google.com/share/SHARE_ID
https://g.co/gemini/share/SHARE_ID
https://share.gemini.google/TOKEN

The first two carry the conversation id. The third is a shortener whose token is not the id, so that form costs one extra request to resolve.

A Gemini conversation may cite web pages and record the searches the model ran. Neither appears in the answer text the provider serves, so both are kept in the JSON archive under provider_metadata, alongside the response ids. As with Claude, no renderer reads provider_metadata and none of it reaches Markdown.

A deleted or unknown Gemini share answers with a success status and an empty payload rather than an error, so Convolvger reports it as a missing share instead of passing an empty conversation off as a real one.

Archive a Claude conversation

A Claude share page cannot be fetched. The page loads its conversation separately, and the service refuses non-browser clients whatever headers they send. Rather than impersonate a browser, Convolvger waits for yours to hand the snapshot over. One-time setup:

convolvger bookmarklet

That prints a bookmarklet and how to save it. Then, for any Claude share link:

convolvger export --capture

Open the share page and click the bookmark. The archive is written and named after the conversation's own title. No URL is typed: it arrives with the snapshot, from the page you clicked. --capture works with inspect too.

Verified in Chrome, Edge and Firefox. Other browsers should work where they run bookmarklets and permit connections to localhost.

If you would rather not run a listener, convolvger export <claude-share-url> prints the URL your browser can read the snapshot from. Open it, save the JSON, and pass it with --from-file. That path works in any browser.

What capturing does, and what it does not

While the command waits it listens on 127.0.0.1:23477 — loopback only, not reachable from your network. It takes one snapshot and then stops, and gives up after three minutes. It accepts a payload only when the browser's origin matches the address the payload claims to come from, so a page on another site cannot post a forged conversation.

Two things it does not defend against, stated plainly: any page on the provider's own domain could reach the listener during the seconds it runs, and so could another program on your machine — though such a program could write the archive file directly anyway. Nothing is transmitted anywhere else. Your browser talks to the provider using the session it already has, and to loopback.

What a Claude archive contains

Claude's snapshot names the account that shared it, so a Claude JSON archive carries created_by and creator — a display name and an account id — under provider_metadata. They are kept because the JSON is the archival record and drops nothing the provider served. No renderer reads provider_metadata, so the Markdown export does not contain them.

Markdown omits provider-hidden messages and deactivated branches by default and reports each omission in its header. --include-hidden and --include-inactive keep them. Both flags are ignored for JSON, which never omits anything.

The header also names what the provider itself did not serve: attachments a message declared but did not carry, and tool results the snapshot referenced and left empty. Observations of the same thing across messages are merged for display and their counts summed, since the message id that separates them in the record is not shown in a Markdown file. Findings about what this tool could not model are counted in the header rather than named, and every finding is recorded in full in the JSON export.

Content whose type this version does not model is flagged rather than dropped, and where such a block carries a title and a link — a web result a search tool returned, for instance — the export shows them. The complete block stays in the JSON. Nothing is fetched to do this: the title and the link were served inside the snapshot, and Convolvger never dereferences a URL it archives.

Check an archive's integrity

convolvger verify mychat.json

Reads a JSON archive and reports what it records, without re-fetching or re-parsing anything. See the Status section below for what the two verdicts mean.

Exit codes

0 clean, 2 completed with warnings or with a failed integrity check, 1 failed.

Goals

• Archive public AI conversations locally
• Support multiple AI providers through isolated adapters
• Preserve conversation structure and exposed content
• Detect incomplete or inconsistent extraction
• Produce deterministic, provider-independent output
• Keep conversion entirely local, with no LLM required
• Avoid account credentials and telemetry

What a snapshot is

Convolvger archives public share snapshots. A snapshot is what a provider chooses to expose at a given URL at a given moment — not the conversation itself, and not a guarantee.

Share pages can change after they are published. While building the ChatGPT adapter we fetched the same share link twice, eight hours apart. Both responses had the same conversation, the same title, and the same 31 message IDs. Four assistant messages addressed to a tool were served the second time with their payload emptied and their content type relabelled from code to text — while their IDs, roles, recipients, statuses, weights and every metadata key stayed identical. A comparison of anything but the content would have called the two snapshots the same file. Nothing about the conversation had changed; the provider's rendering of it had.

That is why every Convolvger export records when it was retrieved, reports every message it could not render, and never silently drops content. An archive is a claim about what was there when you looked. It is only trustworthy if it also says when it looked and what it could not see.

Convolvger does not recover private conversations, hidden model state, chain-of-thought, deleted content, or attachments the provider does not publish. If it is not in the snapshot, it is not in the archive.

Supported Providers

Initial providers:

• ChatGPT — fetched directly from a share link
• Claude — captured from your own browser, since its share pages cannot be fetched
• Gemini — fetched directly from a share link
• Grok — not implemented yet

Additional providers will be added as their public sharing formats are supported and tested.

Status

Convolvger is under active development and not yet ready for general use.

Every command shown under Usage above works today.

ChatGPT share links can be fetched, parsed, and exported. Markdown is a reader-facing document that may omit content, reports every omission in its header, and names what the provider did not serve; JSON is the complete archival record and omits nothing the model holds.

Claude share snapshots can be parsed and exported, but not fetched: the service refuses non-browser clients, so a snapshot is captured from your own browser as described under Usage.

Gemini share links can be fetched, parsed and exported, in all three of the link forms Google issues. Citations and search queries are preserved in the JSON archive rather than rendered into the document. Grok is not implemented yet.

Extraction records structured findings, each with a stable code and a level. A note is recorded but does not change the exit status; a warning does. Both are written into the JSON archive either way, so nothing is withheld from the record because it was judged unremarkable.

convolvger verify PATH reads a JSON archive and reports what it records, without re-fetching or re-parsing anything. It answers two questions separately: whether the provider served everything its own snapshot structure referenced, and whether everything it served could be modelled. Findings that indicate neither are listed in full and counted toward neither answer, so an archive whose messages are mostly empty by design — which is how a public share snapshot normally arrives — is still reported complete. A finding whose code this version does not recognise changes no answer, since an archive may have been written by a later version. An archive declaring a schema version this version cannot read is refused rather than given a verdict.

Exit codes: 0 clean, 2 completed with warnings, 1 failed. For verify, 2 means a check did not pass, and 1 means the file could not be read as an archive at all.

License

MIT