Read Claude.ai Web Conversation

Pull a Claude.ai web conversation into a full, structured transcript — every message and every tool call, not just what is currently on screen.

Verified against Claude.ai's web API as of July 2026. These are the same private JSON endpoints the Claude.ai front-end calls; they are not a documented or version-stable public API, so if a request 404s, re-derive the shape from the Network tab (see references/claude-web-api-extraction.md).

This skill vs. its siblings

Pick by the source you are holding, not by the word "conversation":

Why the obvious approaches fail (read this first)

Four traps, and every one of them fails silently — you get back something that looks like a complete export, and you only notice the loss if you happen to know how long the conversation really was. Assume you are being lied to by default; the fidelity gate in Step 4 exists to make the lies audible.

1. Login wall. curl and WebFetch don't get an auth redirect — they get a Cloudflare challenge page (HTTP 403, Just a moment...). Nothing you do to the headers fixes this; the page is gated on a session that lives in the user's Chrome. Never reach for curl here, not even to "just check".
2. Virtual scrolling. With the conversation open, get_page_text and DOM scraping still only see the handful of messages currently rendered — often just the last one. A 40-message thread comes back as 1.
3. Default rendering collapses the tool calls. The API's default rendering turns every tool call into a "not supported on your current device" placeholder. On a research/agent conversation the tool blocks ARE the content: measured on a real one, the default rendering returned 9.4k chars against 173k with render_all_tools=true — 5%. Always request the full rendering.
4. A /chat/-shaped renderer silently drops /share/ blocks. A share payload returns web_search hits as knowledge items, not text items. A renderer that only knows text returns the empty string for them and reports no error — the same real conversation rendered to 8k chars instead of 146k (4.8% retention) and looked completely fine.

The reliable path: run JavaScript inside the user's logged-in page and let fetch inherit the session cookie, then render locally through the fidelity gate.

Step 0 — Choose an injection channel

Three channels can execute JS inside the user's page. They differ only in plumbing, and each fails in its own way, so pick in this order and verify rather than assume.

⚠️ Expect CDP to be unavailable, and never try to force it. Since Chrome 136 (April 2025) the debugging port is ignored on the default user-data-dir — a deliberate hardening against malware that used CDP to steal cookies. You may still find a listening socket and a stale DevToolsActivePort on disk while the WebSocket handshake is never answered and /json/* returns 404.

And its availability flaps. Verified on Chrome 150, one machine, one browser session: the endpoint worked, then stopped answering entirely, then answered again — with no restart in between. So probe every time and never cache the verdict. "CDP worked five minutes ago" is not evidence that it works now, and "CDP failed once" is not evidence that this machine can't use it. The probe is cheap precisely so you can afford to re-ask.

The trap is what you'll be tempted to do next. Every fix on the web says "relaunch Chrome with --user-data-dir=/tmp/whatever". For this skill that is worse than useless: a fresh profile is signed out, and a signed-out browser cannot read a single one of the user's conversations. The session you need lives in precisely the profile Chrome is refusing to expose. So probe, take the answer, and fall back — never reconfigure or relaunch the user's browser to chase a port.

CDP does work when Chrome was deliberately started on a non-default --user-data-dir and is signed into claude.ai. That happens, and when it does this channel beats the others outright (no menu toggle, immune to the Apple Events capture below, enumerates every tab, returns the whole payload on stdout). That is why it is worth one cheap probe before falling through.

To try channel 1, load the extension's tools in a single ToolSearch call (they are deferred; the API path needs no get_page_text / read_page):

ToolSearch: select:mcp__claude-in-chrome__tabs_context_mcp,mcp__claude-in-chrome__navigate,mcp__claude-in-chrome__javascript_tool

then call list_connected_browsers. An empty list is the account mismatch — go to channel 2 rather than retrying.

To try channel 2:

uv run --with websockets python scripts/cdp_channel.py probe

probe answers three questions at once: is CDP available, is it pointed at the user's real browser (page count, and any claude.ai tabs already open), and are there multiple Chrome instances running. Its output drives everything below — including the AppleScript decision, since the chrome_instances.automation list is what tells you whether Apple Events will even reach the right browser.

Exit codes are a contract, so you never have to guess whether to fall back:

Exit 3 still prints chrome_instances, because that part needs no debugging port. It is a routing fact, not an error: read it, fall through to channel 3, and say nothing to the user about debugging ports.

The AppleScript routing trap (read before you blame the user)

AppleScript addresses Chrome by bundle id. If a second Chrome instance is running — chrome-devtools-mcp, puppeteer, playwright, anything launched with its own --user-data-dir — Apple Events can land on that instance instead, and macOS gives you no way to target the first one by pid. There is no workaround.

What makes this genuinely dangerous is how it presents: the automation profile has "Allow JavaScript from Apple Events" switched off, so your JS attempt fails with "Executing JavaScript through AppleScript is turned off" — and you will conclude the user forgot a menu toggle, and go ask them to flip it, while their real browser had it enabled the whole time. Observed exactly this: AppleScript reported windows=1 tabs=1 while the user's actual Chrome had 35 tabs open.

So before falling back to AppleScript, run probe (or chrome_instances() from cdp_channel.py) and read the automation list. If it is non-empty, say so plainly — "an automation Chrome is holding the Apple Events route, so AppleScript will hit the wrong browser" — and either use CDP, or ask whether that instance can be closed. Do not relay a toggle error you cannot attribute.

Channel details and the base64 binary bridge: references/applescript_fallback_channel.md.

Step 1 — Identify the account case (this decides what you can even get)

How much of the conversation exists for you to fetch depends on whose account the browser is signed into. Establish this early — it is the difference between a complete archive and one with unrecoverable holes, and the user deserves to know which they are getting before you hand it over.

For a /share/… link, the snapshot payload itself tells you which case you're in: it carries conversation_uuid (the original) and creator. So fetch the snapshot first, then try the original with that conversation_uuid — if it 200s you are in case B and should re-export from there, because the snapshot is strictly lossier. If it 404s, you are in case C: say so out loud, and let the render step disclose the gaps in the transcript header (it does this on its own).

Case C's holes are real and worth naming precisely, because they look like a bug in your export when they are not: a view of an uploaded file keeps its name but loses its path and the file's entire content. The user's own copy of that file is usually sitting on their disk — the export isn't broken, it just cannot speak for a file the platform declined to share.

Step 2 — Open the conversation

Extension: tabs_context_mcp with { "createIfEmpty": true }, then navigate. CDP: cdp_channel.py open --url <url> (reuses an existing tab; waits out the Cloudflare interstitial, which the browser clears by itself).

Either way, confirm it actually loaded and is logged in — the tab title should be the conversation's name, not "Log in". If you land on a login page, stop and tell the user; do not automate a login.

Step 3 — Fetch the payload

Always request the full tool rendering. Without render_all_tools=true you get placeholders where the analysis was (trap 3 above).

Note the share endpoint keys on the snapshot id from the URL, which is not the conversation id — feeding a snapshot id to chat_conversations 404s, and that 404 means "wrong id", not "you lack access". Both id families are opaque; always derive them from the open URL rather than hard-coding.

• Extension channel → run scripts/export_conversation.js in the page (fire-and-poll instructions are in its header), then page out window.__claudeExport.rawJson in ~16k .slice() windows.
• CDP channel → run scripts/fetch_cdp.js, which returns the full conversation JSON directly as a Promise, so cdp_channel.py eval resolves it in one call and streams the result to a file:

  uv run --with websockets python scripts/cdp_channel.py eval \
      --match <conversation-or-snapshot-id> --js scripts/fetch_cdp.js --out conversation.json
  

  awaitPromise resolves the async fetch in one call, and the value comes back on stdout, so a multi-megabyte payload never has to cross a context window.
• AppleScript channel → run scripts/export_conversation.js via scripts/runjs.applescript, then poll and read out window.__claudeExport.rawJson (fire-and-poll; see the channel reference).

For a plain chat with no tool calls, where you just need the text right now, the inline snippet in references/claude-web-api-extraction.md assembles a transcript in-page in a single call.

Step 4 — Render locally, through the fidelity gate

uv run python scripts/render_transcript.py conversation.json -o transcript.md \
    --source-url <url>

This handles both payload shapes (/chat/ and /share/), renders every tool call, folds tool outputs into <details>, and collects web_search citations into a "Sources cited" list.

It also refuses to write a lossy transcript. Before emitting anything it audits every block — this block carried N characters; did the renderer emit anything at all for it? — and exits 2 if any block carried text and produced nothing. That is the trap-4 failure, and it is invisible without this check: the markdown looks clean, the exit code is 0, and 95% of the conversation is gone. The budget is measured off the raw payload, deliberately not through the rendering path, because a gate that asks the parser how much there was to render can only ever confirm the parser's own blind spots.

If it fails, it names the offending block and the command to inspect it: teach result_item_text()/render_block() the new shape. Reach for --allow-lossy only when you have consciously decided the loss is acceptable — it is almost never the right answer, and it is never the right first answer.

Blocks the platform emptied (case C) are counted separately as a disclosed gap, never as loss, and are announced in the transcript header. A clean run prints its own accounting:

fidelity: 143,088/143,088 chars rendered (100.0%)
known gap: 12 tool blocks were emptied by the platform (view) — disclosed in the
           transcript header, NOT recoverable from a shared link

The gate audits per item, not per block — items are the granularity content actually gets lost at, and a per-block check credited a whole block as rendered whenever any part of it came out, so a vanished 38k-char item hid behind a 16-char sibling and scored 100%. It also covers what a content[]-only audit structurally cannot see: message-level attachment bodies (a pasted document lives there, not in a block), the top-level text, and messages the active-path walk never reached. And a payload with nothing in it scores 0%, not 100% — an empty fetch is the limit case of the very loss this gate exists for.

If you change the renderer, run the regression suite. Every case in it is a shape that fooled a previous version of the gate:

uv run python scripts/selftest_fidelity.py

Output format and navigation (--toc)

The default output format is Obsidian because Obsidian's Live Preview does not render HTML <details>, so the collapsible tool outputs, thinking, and file bodies would otherwise land as flat noise. The default rewrites each <details> block as a native > [!info]- collapsible callout:

uv run python scripts/render_transcript.py conversation.json -o transcript.md \
    --source-url <url> --toc

Use --format markdown if you need the older HTML <details> output instead.

The conversion runs only after the fidelity gate has passed on the <details> markdown — it is a cosmetic post-pass that touches no payload string, so it can never affect the retention proof (don't move it before the gate). --toc prepends a linked table of contents and inserts per-message anchors (<a id="turn-N">) so long conversations are navigable. --extract-file output is never reformatted.

Other modes, unchanged: --list-files inventories every downloadable file with the endpoint family each needs; --extract-file <sandbox-path> reconstructs a sandbox-created file by replaying its create_file plus every later str_replace, and refuses to emit anything if a replacement is missing or ambiguous rather than handing back stale content.

Completion check — render is not done while files remain

render_transcript.py prints two lines to stderr when it writes a transcript. Read both before declaring the render finished:

fidelity: 143,088/143,088 chars rendered (100.0%) across 7 message(s)
files: 2 downloadable file(s) in this conversation — the transcript names them but does NOT carry the bytes, so a text-only export is an INCOMPLETE archive. ...

The fidelity: line is the retention proof. The files: line is the completion signal: if it reports any files, the transcript references them by name but does not contain their bytes — a text-only export is an incomplete archive. The whole reason this line lives on the same stderr stream as fidelity: is that completion-drive otherwise sees 100% and stops before the files are pulled, which is exactly how a text export gets mistaken for a full archive. (No files: line at all means the conversation carries no downloadable files; the render is genuinely complete on its own.)

So when files: reports N > 0, downloading them is part of this step, not an optional follow-up. A user who said "archive / export / 拉到本地" asked for a complete archive; handing them the transcript while the files it references sit un-downloaded is handing them an incomplete one. (--list-files / --extract-file never print this notice — they return before the render branch.)

Use scripts/download_files.js to inventory and pull every upload, assistant image, and sandbox deliverable through the correct endpoint family. The script is channel-agnostic JS; run it through the same channel you fetched with. CDP (one call, result on stdout):

# 1. Fire the inventory/download
uv run --with websockets python scripts/cdp_channel.py eval \
    --match <conversation-or-snapshot-id> --js scripts/download_files.js
# 2. Poll until the status line is no longer 'pending' (a few seconds)
echo "window.__dlStatus" > /tmp/check_dl.js
uv run --with websockets python scripts/cdp_channel.py eval \
    --match <conversation-or-snapshot-id> --js /tmp/check_dl.js
# 3. Extract one file at a time (repeat for each name the status line listed)
echo "window.__dl['filename.png']" > /tmp/read_one.js
uv run --with websockets python scripts/cdp_channel.py eval \
    --match <conversation-or-snapshot-id> --js /tmp/read_one.js 2>/dev/null \
  | tr -d '\n' | base64 -d > filename.png

AppleScript channel — the realistic fallback when CDP is unavailable (Step 0 says CDP is usually ignored on the default profile), so it gets a worked example too, not just a swap pointer:

# 1. Fire the inventory/download
osascript scripts/runjs.applescript scripts/download_files.js <conversation-id>
# 2. Poll until the status line is no longer 'pending' (a few seconds)
echo "window.__dlStatus" > /tmp/check_dl.js
osascript scripts/runjs.applescript /tmp/check_dl.js <conversation-id>
# 3. Extract one file at a time (repeat for each name the status line listed)
echo "window.__dl['filename.png']" > /tmp/read_one.js
osascript scripts/runjs.applescript /tmp/read_one.js <conversation-id> \
  2>/dev/null | tr -d '\n' | base64 -d > filename.png

Extension channel: run the script body via javascript_tool and page window.__dl out in ~16k slices, like any other large return.

Verify each file with file <name> (magic bytes — expect PNG image data, Microsoft Excel 2007+, not ASCII text) and compare its byte size against the metadata in the conversation JSON (files[].size_bytes for uploads; the status line from download_files.js for deliverables). Save downloads next to the transcript — conventionally a sibling <transcript-basename>_附件/ directory — and repoint the transcript's image references at the local copies.

Step 5 — Paging (extension channel only)

javascript_tool truncates large return values. Fetch the chars count first, then re-run returning later windows — keep the fetch identical and change only the final expression:

transcript.slice(14000, 32000);   // then (32000, 50000) … until you've covered `chars`

Prefer ~14–18k windows; larger risks hitting the limit again. The CDP and AppleScript channels don't need any of this — both return the whole payload on stdout, which is a good reason to prefer them for a large archival export.

Gotchas

• sender values are 'human' and 'assistant' (not 'user'/'claude').
• A message can have m.text AND m.content[] at the same time. Agent turns often carry the final answer in m.text plus thinking/tool_use/tool_result blocks in content[]. Build from content[] first and fold in m.text — never m.text || (content…), which short-circuits and drops every block whenever m.text is set. If a message renders blank, inspect one raw: Object.keys(msgs[0]) and msgs[0].content?.map(b => b.type).
• A /share/ payload is shaped differently from /chat/. name is null (the title is in snapshot_name); web_search results are knowledge items (title/url/text), not text items; and citations hang off the text blocks. The bundled renderer handles all three — a hand-rolled one usually doesn't, which is trap 4.
• Empty input: {} / content: [] on a tool block is not a bug — on a share snapshot it is the platform withholding the sharer's private file contents. Say so; don't render it as a mysterious no-op, and don't count it as data loss.
• If rendering_mode=raw returns empty or short bodies, retry with rendering_mode=messages. The two expose slightly different fields.
• The conversation's files are downloadable too — uploads, and the deliverables behind Download cards. Two endpoint families, keyed differently (images by uuid; uploads/outputs by sandbox path via conversations/<id>/wiggle/download-file). uuid-guessing 404s for uploads, which looks like — but is not — "unsupported". scripts/download_files.js inventories the conversation and pulls each through the right endpoint. Don't conclude anything is unavailable before trying it.
• tree=True returns the whole tree, including branches abandoned by edits and regenerations. Walk the active path from current_leaf_message_uuid via parent_message_uuid so dead branches don't leak into the transcript or inflate the message count. The renderer does this and falls back to array order when those fields are absent (correct for single-chain conversations, and for share snapshots, which are already linear).
• Multiple organizations: orgs[0] may be the wrong one. If a conversation 404s, list them — orgs.map(o => ({uuid: o.uuid, name: o.name})) — and loop the fetch across orgs before concluding it's inaccessible.
• Just read — don't click. This skill never needs to touch the conversation UI; avoid triggering navigation or dialogs mid-fetch.
• Never relaunch the user's browser to open a debugging port. Beyond the usual reason (it's their browser, not yours), it is self-defeating here: Chrome only honours the port on a non-default profile, and a fresh profile is signed out — so the browser you just launched cannot read any of their conversations. You would be trading the session you need for a port you don't. An unavailable port is a fact to route around, not a setting to change.

Full endpoint table, response schema field-by-field, the complete export script, and a troubleshooting table: references/claude-web-api-extraction.md.

Next Step

Once the transcript is written and its files are downloaded — the files: stderr line from Step 4's Completion check is what tells you whether there were any, and a text-only export stays incomplete while that count is non-zero — suggest the natural follow-up. Opt-in, never automatic:

Got the full conversation (<N> messages, "<title>", <retention>% fidelity);
<N> files downloaded to <dir>.

Options:
A) Clean it up — run transcript-fixer if it's ASR/garbled (only if relevant)
B) Summarize / extract the decisions and action items
C) Save it to a file — tell me where
D) Nothing else — you just needed it read