Claude Desktop Chat Export

claude-desktop-chat-export

claude.ai chats aren't resumable — once you hit context limit or close the tab, the conversation is read-only. Claude Code CLI stores every session as JSONL at ~/.claude/projects/<encoded-cwd>/<sessionId>.jsonl and resumes via claude -r <sessionId>. This skill bridges the two: export a claude.ai conversation, convert it to CC's format, keep talking.

Same underlying model. Only the harness differs (system prompt, declared tools, output-length rules).

Workflow

Three stages: grab from the browser → organize → convert. The converter is scripts/convert_to_cli.py.

Two transport options for the "grab" stage:

• DevTools console — downloads conversation.json and chat-images.zip to ~/Downloads via the browser's download mechanism. Works when the user runs the snippets manually.
• MCP + local relay — works when driving the browser via the claude-in-chrome MCP. The MCP's javascript_tool return filters block cookie-like and base64 payloads, and extension-driven a.click() downloads don't land on disk. Solution: run scripts/relay_server.py locally, have the browser POST each blob to it. See MCP path below.

1. Grab the conversation JSON + images from claude.ai

Requires an authenticated browser session at claude.ai. Use DevTools console or the claude-in-chrome MCP.

First, identify the org and conversation UUIDs:

• Org UUID: from /api/bootstrap → memberships[0].organization.uuid
• Conversation UUID: from the chat URL

Run in the browser console (on claude.ai, already logged in):

const ORG = '<your-org-uuid>';
const CONV = '<conversation-uuid>';

// Fetch full conversation
fetch(`/api/organizations/${ORG}/chat_conversations/${CONV}?tree=True&rendering_mode=messages&render_all_tools=true`,
      {credentials:'include'})
  .then(r => r.blob())
  .then(b => { const a = document.createElement('a'); a.href = URL.createObjectURL(b); a.download = 'conversation.json'; a.click(); });

For images (API only serves webp previews — no originals):

// Load JSZip once:
await new Promise((res,rej)=>{ const s=document.createElement('script'); s.src='https://cdnjs.cloudflare.com/ajax/libs/jszip/3.10.1/jszip.min.js'; s.onload=res; s.onerror=rej; document.head.appendChild(s); });

// Fetch + zip all images:
const convo = await (await fetch(`/api/organizations/${ORG}/chat_conversations/${CONV}?tree=True&rendering_mode=messages&render_all_tools=true`, {credentials:'include'})).json();
const images = [];
for (const m of convo.chat_messages) for (const f of (m.files||[])) if (f.file_kind==='image') images.push(f);
const zip = new JSZip();
for (const img of images) {
  const r = await fetch(`/api/${ORG}/files/${img.file_uuid}/preview`, {credentials:'include'});
  if (r.ok) zip.file(`${img.file_uuid}.webp`, await r.arrayBuffer());
}
const blob = await zip.generateAsync({type:'blob'});
const a = document.createElement('a'); a.href = URL.createObjectURL(blob); a.download = 'chat-images.zip'; a.click();

Chrome blocks a second auto-download from the same page — either run the two snippets in separate fresh tabs, or click "Allow" on the download-permission prompt.

MCP path (via relay)

When driving the browser via the claude-in-chrome MCP, use the bundled relay. Start it pointing at the target export directory:

python3 <skill-dir>/scripts/relay_server.py <export-dir>/<new-uuid>/ &

It listens on http://127.0.0.1:8765 by default and writes each POST body to <out-dir>/<X-Filename>. Nested paths like files/<uuid>.webp are preserved; .. / absolute paths are rejected.

Then, via javascript_tool on a claude.ai tab, POST the conversation JSON and each image blob:

// Conversation JSON
(async () => {
  const ORG = '<org-uuid>';
  const CONV = '<conversation-uuid>';
  const r = await fetch(`/api/organizations/${ORG}/chat_conversations/${CONV}?tree=True&rendering_mode=messages&render_all_tools=true`, {credentials:'include'});
  const blob = await r.blob();
  const up = await fetch('http://127.0.0.1:8765/upload', {
    method:'POST',
    headers:{'Content-Type':'application/json','X-Filename':'conversation.json'},
    body: blob
  });
  return {size: blob.size, relay: up.status};
})()

// All image previews, written to files/<uuid>.webp
(async () => {
  const ORG = '<org-uuid>';
  const CONV = '<conversation-uuid>';
  const convo = await (await fetch(`/api/organizations/${ORG}/chat_conversations/${CONV}?tree=True&rendering_mode=messages&render_all_tools=true`, {credentials:'include'})).json();
  const imgs = [];
  for (const m of convo.chat_messages) for (const f of (m.files||[])) if (f.file_kind==='image') imgs.push(f.file_uuid);
  let ok=0, fail=0;
  for (const uuid of imgs) {
    try {
      const r = await fetch(`/api/${ORG}/files/${uuid}/preview`, {credentials:'include'});
      if (!r.ok) { fail++; continue; }
      const up = await fetch('http://127.0.0.1:8765/upload', {
        method:'POST',
        headers:{'Content-Type':'application/octet-stream','X-Filename':`files/${uuid}.webp`},
        body: await r.blob()
      });
      if (up.ok) ok++; else fail++;
    } catch(e) { fail++; }
  }
  return {total: imgs.length, saved: ok, failed: fail};
})()

Stop the relay (kill %1 / Ctrl+C) once both uploads are done. This bypasses both the MCP's cookie-data filter (only metadata, not page data, is returned from JS) and the extension-initiated-download issue.

2. Organize

mkdir -p <export-dir>/<conversation-uuid>/files
mv ~/Downloads/conversation.json <export-dir>/<conversation-uuid>/conversation.json
unzip ~/Downloads/chat-images.zip -d <export-dir>/<conversation-uuid>/files/

Result:

<export-dir>/<conversation-uuid>/
├── conversation.json      ← full claude.ai API response
└── files/
    └── <file_uuid>.webp   ← one per attached image

(With the MCP + relay path, files land directly in this layout — no unzip needed.)

3. Convert

python3 <skill-dir>/scripts/convert_to_cli.py \
  <export-dir>/<conversation-uuid>/conversation.json \
  --cwd <project-path> \
  --session-id <conversation-uuid> \
  --slug <any-slug-string>

Output lands at ~/.claude/projects/<encoded-cwd>/<conversation-uuid>.jsonl. Resume:

claude -r <conversation-uuid>

No external Python dependencies — stdlib only (Python 3.10+).

CLI flags

Defaults to know

Tools flattened by default. claude.ai tool names (view, web_fetch, bash_tool, …) collide with CC's declared tool names (Read, Bash, …). Resuming with raw tool blocks produces API errors. The default --keep-tools off flattens them to readable text so the model retains context of what was looked up. Only pass --keep-tools if you know what you're doing.

Thinking blocks dropped by default. Their cryptographic signatures are tied to the original API context and don't transfer, so keeping them as structural thinking blocks triggers "Invalid signature in thinking block" on resume. Pass --flatten-thinking to preserve thinking as plain <thinking>…</thinking> text blocks — no signatures, no resume failures, model reads it as ordinary context. Later, session-stripper strip-thinking detects these wrappers and can clear them without re-converting. Otherwise, the raw thinking text is still readable in conversation.json.

UUIDs regenerate on every converter run. Only sessionId is stable (via --session-id). If a CC session for this conversation is currently open, quit CC before re-running the converter — otherwise CC's in-memory state points to deleted UUIDs and the chain breaks silently.

For the full list of format-level gotchas, chain integrity requirements, and what does/doesn't survive the import, see references/format-notes.md.

When something goes wrong

• "tool use concurrency issues" on resume: ran with --keep-tools but tool names don't match CC's declarations. Re-run without --keep-tools.
• "Invalid signature in thinking block": thinking blocks leaked through. Re-run; converter drops them by default.
• "Extra inputs are not permitted": tool_result had stray fields (uuid, citations, start_timestamp). Converter whitelists valid keys — re-run.
• Chain broken silently / empty session list: CC was open during the previous conversion. Quit CC, re-run.
• Images missing in resumed chat: check --files-dir points to a folder with <file_uuid>.webp files matching the JSON's files[].file_uuid.

See references/format-notes.md for the full gotcha list and the chain integrity invariants (parentUuid unbroken, slug identical, timestamps monotonic).