Browser conversation capture
Hansard Browser is a preview Chromium extension published in the Chrome Web Store. Its published release, 0.6.1.0, captures ChatGPT, Claude.ai, Gemini, and Perplexity conversations, plus Devin cloud sessions and Ask Devin queries. A later store release adds every site the source build captures. It is separate from the npm package.
The current source build adds experimental capture for Claude Duet, Claude Code on the web, Codex cloud tasks, ChatGPT Work Cloud and dots, the workspace a Claude.ai chat links, Claude Design, Perplexity Computer, Gemini Notebook / NotebookLM, Muse (Personal Agent), and Grok, Meta AI, DeepSeek, Kimi, Qwen Chat, Mistral chat (Le Chat), Poe, and Z.ai conversations. The Chrome Web Store release does not include these yet. Exact product coverage marks each source-build product. Update the extension and Hansard together, since the receiver in the current Hansard release imports their captures.
Install and pair
Section titled “Install and pair”Install Hansard from the Chrome Web Store in Chrome, Brave, Edge, or another Chromium browser (version 109 or later). In Edge, first allow extensions from other stores. Guided setup’s Web chats guide, also under Guided setup in Settings → Imports, walks through each step and draws what the store, the browser, and the extension show:
- On the listing, select Add to Chrome, then Add extension. Google’s guide covers installing and pinning.
- Select the puzzle icon in the toolbar, then the pin beside Hansard.
- Pair the browser with the code Hansard shows, as below.
- On the extension’s page, switch on each site under Cloud chats. The browser asks once for access to each site; select Allow.
- Keep a signed-in tab open for each site. The guide’s last step shows the first chat as it arrives.
Pair in the app
Section titled “Pair in the app”- Open Settings → Imports and find Browser extension.
- Next to Connect a browser, select Pair…. Open Chrome Web Store opens the listing. In a source checkout, Show development build reveals the unpacked build to load instead.
- Paste the code from Paste this code in the extension’s settings into the extension’s settings. The code works once and expires after ten minutes.
Each paired browser gets its own row with Disconnect. The Web chats row shows when a web chat was last saved and, from extension 0.7.1.2, what each provider waits for, such as an open tab or another account. When no web chat has been saved for a day and the extension names a reason, the title band shows it too. The Captures row counts saved, waiting, unchanged, and failed captures, names the most common reason for failures, and offers Retry. While the app is open it keeps the receiver running for a paired browser, as the watcher does.
Pair from the CLI
Section titled “Pair from the CLI”hansard browser serveIn another terminal, run:
hansard browser pairPaste the pairing code into the extension settings. hansard browser pair
runs only from your own terminal, never from an agent session.
hansard browser status
lists paired browsers and their captures, hansard browser retry retries
failed imports now, and hansard browser revoke <installation-id> disconnects
a browser. The watcher keeps the paired receiver running; otherwise keep
hansard browser serve open.
A receiver started from the CLI has priority. After it exits, the watcher takes over on the last port while a browser is paired or a pairing code is valid. Captures remain queued across browser restarts when no receiver runs.
Before the extension sends a pairing code, its credential, or a capture, the
receiver proves it holds that pairing, so another program listening on the
receiver’s port receives nothing and captures wait in the browser. A browser
paired by an earlier release may be asked to pair again. If another program
holds the port, the receiver does not start, and hansard browser status and
the Browser extension settings name that program when the system reports
it. Stop it, or run hansard browser serve --port with a free port and set the
extension’s Hansard address to match.
Your browser updates the extension on its own after the Chrome Web Store publishes a release. Hansard accepts every version of the extension, so an update waiting on store review never stops captures. If Hansard is too old to prove itself to the extension, the extension asks you to update Hansard. A capture from a provider your Hansard cannot import yet waits in the queue, counted as waiting for a Hansard update, and imports after you update.
Build from source
Section titled “Build from source”To load a development build instead of the store release, build it from a Hansard source checkout:
npm --prefix extension cinpm --prefix extension run buildOpen the browser’s extensions page, enable Developer mode, choose Load
unpacked, and select extension/dist.
Capture history
Section titled “Capture history”Enable the providers you want in the extension’s settings. Each provider requests its own website and file-host permissions, and the settings show each provider’s status. Keep a signed-in tab open for that provider. Choose Sync now to capture history or Rescan all history to refresh existing captures. The toolbar icon shows a badge when captures fail. Failed captures stay queued and retry on their own, waiting longer after each failure, up to six hours.
Until the Chrome Web Store publishes them, these changes need the source build:
- Each provider’s status lists delayed retries apart from ready conversations, with the reason and a countdown for the next retry. One conversation’s failure does not stop the others unless the provider asks for a pause.
- When a provider is signed in to another account, the earlier account’s conversations wait, named in the extension’s settings, and capture continues for the account signed in now. They resume once that account is signed in again.
- A ChatGPT reply that stops changing for ten minutes is saved as it stands and saved again when it updates. A reply still generating never holds up other conversations.
- A conversation that keeps failing for a day while others save is set aside until it changes or you choose Rescan all history.
- The Claude.ai scan also lists archived conversations and conversations from Claude’s unified history.
- The extension captures the tab whose conversation changed first, and current conversations take turns with older history.
- ChatGPT collection slows down after repeated rate limits.
Hansard downloads the public images and code charts that Perplexity captures reference, up to 100 MiB per file and 500 MiB per conversation. A host that asks for a pause holds back only its own downloads, and the receiver takes turns between providers so one provider’s queue does not hold up another’s. Rescan all history also downloads the images of earlier Perplexity captures.
Imported and unchanged captures leave the receiver’s queue after a week.
Their totals remain visible in the extension, Settings, and
hansard browser status. Perplexity downloads keep durable file receipts and
use no browser cookies. The limits also include 500 attachments per
conversation. Missing bytes stay explicitly unavailable, and a rescan cannot
recover files the provider no longer serves.
Where captured chats appear
Section titled “Where captured chats appear”Every captured chat service appears in Sessions as a web chat, marked with the service’s logo. When you set a picture in Settings → General → Your picture, it stands beside the account’s name. Chats that belong to no project go in a folder named for the service followed by not in a project, such as ChatGPT, not in a project. Select a folder’s name to open its chat page. Dots appears as its own app beside ChatGPT. To connect a chat account to an account in Accounts, use Link chat account… there. See Accounts and plan usage.
Rebuild captures without a browser
Section titled “Rebuild captures without a browser”Hansard keeps the provider responses behind each browser capture. To rebuild those conversations from the archive with the current importer, without opening a browser, run the command for that provider:
hansard rebuild --from-raw --provider chatgpthansard rebuild --from-raw --provider perplexityhansard rebuild --from-raw --provider claude_webhansard rebuild --from-raw --source-type devin-cloud-browserhansard rebuild --from-raw --source-type devin-ask-browserhansard rebuild --from-raw --source-type muse-agent-exporthansard rebuild --from-raw --source-type codex-cloud-browser --since allhansard rebuild --from-raw --source-type perplexity-computer-browser --since allhansard rebuild --from-raw --source-type notebooklm-browser --since allhansard rebuild --from-raw --source-type gemini-web-browser --since allAfter an update, the watcher’s refresh after updates parses Codex cloud and Perplexity Computer captures again on its own, so Codex cloud tasks take their pull requests’ repository and Perplexity Computer replies show. Run the last two commands only when automatic refresh is off.
The ChatGPT rebuild, which includes Work Cloud and dots, restores project membership and identifies dots replies. A Claude.ai conversation that captures and exports both shaped replays from the order its import recorded; one archived before Hansard kept that record refuses replay and keeps its archive. Replay cannot download files the archive lacks, and older captures cannot recover details whose responses were not kept.
Before continuing a Dots conversation from an older archive, refresh its recorded reply roles:
hansard rebuild --from-raw --source-type chatgpt-dots-browserClaude.ai replay retains the latest browser capture, up to 32 earlier captures, eight necessary exports, and 128 MiB of earlier capture data. An input retires only when the remaining inputs reproduce the same conversation. These replay records are Hansard metadata, separate from preserved provider responses.
An older Claude.ai browser history without a replay record keeps its archive
and refuses replay. A new import becomes replayable only when its inputs
reproduce all retained messages, metadata, and files. For a fresh replayable
copy, import the account export and capture the conversation into a separate,
empty archive with another HANSARD_HOME. Keep the original archive, since
current provider inputs can lack its earlier files or metadata. Hansard does
not migrate or delete it automatically. Legacy conversation-style memory
exports cannot replay. Current Claude.ai export memories remain separate
read-only records in Memories.
See What the provider reports for conversation settings, projects, and alternate branches. Captures use their capture time. Exports use their recorded export time, or the conversation’s update time when no export time exists. Provider times without a time zone remain as reported and are not treated as instants.
Exact product coverage
Section titled “Exact product coverage”| Product | Archived format | Preserved data and limits |
|---|---|---|
| ChatGPT | chatgpt-browser |
History, archived chats, project conversations, conversation trees, the active branch, exposed model and file metadata, and referenced canvas documents. The source build keeps generated files that messages link, slows collection after repeated rate limits, keeps its place through them, and retries individual collection failures independently. Updated-extension recovery and project, GPT, and Voice field comparisons still need native acceptance. Unavailable branches and expired media remain gaps. |
| ChatGPT Work Cloud | chatgpt-browser |
Experimental source build: active and archived Work task lists, full conversation trees, and a separate Work Cloud label. Native account inventory, output files, and automatic resume remain unverified. |
| Dots | chatgpt-dots-browser |
Experimental source build: paginated ChatGPT messaging-room history, recorded models, attachment descriptors, permitted file downloads, and separate task links. Other channels, memory, voice, and delegated task transcripts remain uncovered. Native message and attachment acceptance remains pending. |
| Codex cloud | codex-cloud-browser |
Experimental source build: current and archived tasks, complete turn mappings, individual turn events, recorded models and usage, and raw alternate attempts. Capture requires the signed-in workspace identity and supports at most 500 turns per task. A task is filed under the repository its recorded GitHub pull requests name, when they name exactly one; an environment label alone assigns no repository. Task-file downloads and native acceptance remain open. |
| Claude.ai | claude-web-browser |
Organization conversation trees, exposed projects and settings, uploaded and generated files; preview bytes remain labeled as previews. |
| Claude.ai linked workspaces | claude-cowork-cloud-history |
Experimental source build: explicitly linked workspace events and tools archive separately from the chat. Recorded Write bytes and current remote files stay labeled by their source. File acquisition follows recorded paths within size limits. Stops, missing files, and unread pages remain explicit. Native extension acquisition remains unverified. |
| Claude Duet | claude-duet-history |
Experimental source build: Duet’s channels at claude.ai/duet, each task as its own conversation linked to its channel’s main thread, with its status, replies, approvals, Make outputs that Claude returns with their contents, and the work of its cloud sessions. Turn on Claude Duet in the extension’s settings. Signed-in validation remains pending. |
| Claude Code on the web | claude-cloud-history |
Experimental source build: claude.ai/code sessions, including archived ones, with every retained event, tool calls, recorded reply models and usage, and the session’s repository. A session also read by hansard import claude-cloud stays one conversation. A capture whose last event page says more events follow waits as a failed capture until Rescan all history reads it in full. Environment variable values are redacted. Session files and sessions over the 32 MB capture limit are not captured. Signed-in validation remains pending. |
| Claude Design | claude-design-browser |
Experimental source build, under the Claude provider: open claude.ai/design in a signed-in tab to capture standalone Design project chats, current files, and exposed revisions, up to 8 MiB per file and 64 MiB per project. Account-wide project inventory, direct edits, and closed chats remain unverified. Standalone Design closes December 14, 2026. |
| Gemini chat | gemini-web-browser |
Paginated history, selected response candidates, typed media, and recorded response models. Captured Canvas documents and code enter the Library with their observed versions. Canvas history coverage is experimental and signed-in payload verification remains pending. Takeout does not supply missing Canvas content. Media coverage is partial. |
| Gemini Notebook (NotebookLM) | notebooklm-browser |
Experimental source build: settled messages visible in an open notebook, expanded Thoughts, citation labels, and exposed source IDs. Open Studio notes can enter the archive and Library, but this capture remains unverified on a live notebook. Native chat continuation reaches the receiver and passes isolated installed import and replay. Production archive and viewer acceptance remain pending. Unopened notes, audio overviews, mind maps, citation-to-source identity, and account-wide history are not captured. |
| Perplexity | perplexity-browser |
Search threads, citations, research steps, recorded models, and downloaded public images and code charts, up to 100 MiB per file and 500 MiB per conversation. Previews and unavailable files remain explicit. |
| Perplexity Computer | perplexity-computer-browser |
Experimental source build: replies, user messages, thinking, retained workflow text, tools, subagent identities, statuses, and exposed artifacts. Account-wide memory, Skills, connectors, and automation definitions remain outside capture. Native acceptance is open. |
| Grok chat | grok-chat-browser |
Experimental source build: saved and workspace chats with full response trees, reasoning, recorded response models, and tool results. Complete account inventory, original media, workspace files, tasks, and Imagine and Companion history are not captured. Grok Build and Grok Bot are separate products. |
| Meta AI | meta-ai-browser |
Experimental source build: pinned and unpinned chats with message branches and recorded reasoning. Tool details and citations stay in the raw capture. Original files and media, device and messaging history, memory, and account exports are not captured. Signed-in validation remains pending. |
| DeepSeek chat | deepseek-chat-browser |
Experimental source build: saved and pinned chats with alternate messages, reasoning, file details, and search and open tool calls. Original files, images, voice, and official exports are not captured. Signed-in validation remains pending. |
| Kimi chat | kimi-chat-browser |
Experimental source build: global, pinned, and project chats with reasoning and paged tool calls and results. Original files, generated artifact revisions, project resources, separate tasks, and hidden branches are not captured. The current model selector is not read as the model a reply used. |
| Qwen Chat | qwen-chat-browser |
Experimental source build: normal, pinned, archived, and project chats with the selected branch, reasoning, and recorded tool calls. Hidden branches, temporary chats, original files, and project instructions are not captured. Signed-in validation remains pending. |
| Mistral chat (Le Chat) | mistral-chat-browser |
Experimental source build: private, archived, shared, pinned, and project chats in the selected workspace, with message versions, canvas, and linked research. Original files, workflow runs, project resources, Libraries, and official exports are not captured. Signed-in validation remains pending. |
| Poe | poe-browser |
Experimental source build: saved multi-bot chats with canvas content and member lists, without guessed models. Original resources, earlier edit versions, bot instructions, and official exports are not captured. Temporary and deleted chats may not be recoverable. Signed-in validation remains pending. |
| Z.ai chat | zai-chat-browser |
Experimental source build: normal, pinned, and archived chats with the active branch, alternate branches, reasoning, recorded models, and tool calls. Original files, workspaces, subagents, scheduled tasks, and account exports are not captured. Signed-in validation remains pending. |
| Devin cloud | devin-cloud-browser |
Cloud event history, tools, child-session relationships, exposed files and metadata; private signed-storage payloads may remain unavailable. |
| Ask Devin | devin-ask-browser |
Questions, reasoning, codebase searches, citations and recorded model; missing per-question times remain unknown. |
| Muse (Personal Agent) | muse-agent-export |
Experimental source build: requests Muse’s agent data export through the open, signed-in main chat and transfers it in resumable chunks, up to the 256 MiB archive limit. It also reads side chats with their tool calls and output, goals, task runs, approvals, and the original MEMORY.md. Update the extension and Hansard together. Models, usage, and history completeness are unavailable. Signed-in validation remains pending. |
The extension does not capture Grok Bot, which remains in the coverage backlog.
Most collectors use private website APIs. NotebookLM reads visible page content. Unsupported response shapes pause capture with an error. A closed, unloaded, suspended, or signed-out tab cannot supply history. Deleted and temporary chats, project-only lists, unavailable branches, and expired files are not guaranteed to be recoverable.
Overlapping routes
Section titled “Overlapping routes”Manual ChatGPT, Claude.ai, and Gemini exports remain supported. Captures and exports can enrich the same account conversation when their provider account and conversation identities agree. Matching titles or account labels are not enough to merge them. Devin desktop mirrors, API records, and browser captures likewise retain verified cloud identity while preserving the most complete transcript. See the individual import guides for each route’s limitations.
After capture, open the conversation in Hansard and check messages, files, source, account, project, models, and the Details panel. A successful delivery receipt alone does not prove that every provider field was preserved.