Skip to content

CLAUDE_CONVERSATIONS_API.md

Latest commit

 

History

History
180 lines (127 loc) · 5.15 KB

CLAUDE_CONVERSATIONS_API.md

File metadata and controls

180 lines (127 loc) · 5.15 KB
Raw

Claude.ai Conversation History API — Reverse Engineering Notes

Date: 2026-06-19
Status: ✅ Confirmed via Chrome DevTools (CDP) on signed-in session
Method: Attached to user's Chrome tab (remote-debugging-port=9222)

Summary

Claude web history uses claude.ai session cookies (not Paprwork Claude OAuth tokens). The OAuth scopes used for inference (user:profile, user:inference) do not grant access to these endpoints.

This is not parallel to ChatGPT's backend-api/conversations + OAuth Bearer flow.

Confirmed Endpoints

1. List conversations (sidebar)

GET https://claude.ai/api/organizations/{org_uuid}/chat_conversations_v2
  ?limit=30
  &starred=false
  &consistency=eventual

Notes:

  • chat_conversations_v2 (not the older chat_conversations path some tools document)
  • Pagination via limit + likely offset (offset observed in older chat_conversations docs)
  • consistency=eventual appears required for list reads

2. Single conversation (full messages)

GET https://claude.ai/api/organizations/{org_uuid}/chat_conversations/{conversation_uuid}
  ?tree=True
  &rendering_mode=messages
  &render_all_tools=true
  &consistency=strong

Response (confirmed shape):

{
  "uuid": "e7f85eaf-bd80-4b32-a847-deb9f7f32909",
  "name": "Pre-seed data room checklist for AI infra companies",
  "summary": "",
  "model": "claude-opus-4-5-20251101",
  "created_at": "2026-01-19T19:13:07.864662Z",
  "updated_at": "2026-01-21T01:10:11.548996Z",
  "settings": { ... },
  "is_starred": false,
  "is_temporary": false,
  "platform": "CLAUDE_AI",
  "chat_messages": [ ... ]
}

Messages use tree structure (parent_message_uuid, current_leaf_message_uuid on conversation).

3. Bootstrap / org discovery

GET https://claude.ai/edge-api/bootstrap/{org_uuid}/app_start
  ?statsig_hashing_algorithm=djb2
  &growthbook_format=sdk
  &include_system_prompts=false

Returns account, org uuid, model configs, memberships (~400KB). Useful to discover org_uuid if unknown.

GET https://claude.ai/api/bootstrap/{org_uuid}/current_user_access

Returns feature flags and permissions (includes export:data).

Authentication

What works (web client)

Mechanism Details
Cookie sessionKey=sk-ant-sid01-... (HttpOnly, set on login)
Client headers See below

Required / observed headers (no Authorization: Bearer):

anthropic-client-platform: web_claude_ai
anthropic-client-version: 1.0.0
anthropic-client-sha: <git sha of web bundle>
anthropic-anonymous-id: claudeai.v1.<uuid>
anthropic-device-id: <uuid>
x-activity-session-id: <uuid>
content-type: application/json
User-Agent: Mozilla/5.0 ...
Referer: https://claude.ai/chats

Optional telemetry: x-datadog-*, traceparent, tracestate

What does NOT work

  • Paprwork Claude OAuth access token (sk-ant-oat01-...) → inference API only
  • API keys (sk-ant-api03-...) → Platform API, not claude.ai web

How We Captured This

Prerequisites

Chrome with remote debugging (user already had port 9222 open):

# If needed:
/Applications/Google\ Chrome.app/Contents/MacOS/Google\ Chrome --remote-debugging-port=9222

Probe script (attach to signed-in tab)

node scripts/probe-claude-existing-tab.mjs

Opens /chats, clicks first chat, logs Network via CDP.

Manual DevTools steps

  1. Open claude.ai in Chrome (signed in)
  2. DevTools → Network → filter chat_conversation
  3. Go to Chats (/chats) → see chat_conversations_v2
  4. Open any chat → see chat_conversations/{uuid}?tree=True...
  5. Right-click request → Copy as cURL (includes cookies)

Comparison: ChatGPT vs Claude

ChatGPT Claude.ai
Auth OAuth Bearer token Session cookie sessionKey
List /backend-api/conversations /api/organizations/{org}/chat_conversations_v2
Detail /backend-api/conversation/{id} /api/organizations/{org}/chat_conversations/{id}
Paprwork OAuth ✅ Works ❌ Different auth surface
Official import N/A Settings → Privacy → Export data (ZIP)

Implementation Options for Paprwork

Option A: Official ZIP import (recommended)

User exports from claude.ai → upload conversations.json → parse into Paprwork chats. ToS-safe, no session hijacking.

Option B: Session cookie + unofficial API

Like claude-explorer: Playwright login → capture sessionKey → call endpoints above. Fragile (Cloudflare, expiry, unofficial).

Option C: OAuth like ChatGPT

Not viable — scopes and token type do not cover web conversation APIs.

Probe Scripts

Script Purpose
scripts/probe-claude-existing-tab.mjs Attach to signed-in Chrome tab
scripts/probe-claude-history-cdp.mjs Open new tab via CDP
scripts/probe-claude-conversation-detail.mjs Single chat page capture

Related Docs

  • docs/CHATGPT_CONVERSATIONS_API.md — ChatGPT OAuth approach (works differently)
  • docs/CHATGPT_CONV_HISTORY_IMPLEMENTATION.md — Paprwork ChatGPT UI