OpenAdapter - Claude.ai to OpenAI API Bridge for OpenClaw

Self-hosted OpenAI-compatible API for Claude.ai using Playwright automation. Perfect for OpenClaw, custom AI assistants, and any tool expecting OpenAI's API format.

🦞 Built with OpenClaw in mind — Use Claude's intelligence without API costs!

No API key required • Free & Open Source • Streaming Support • File Uploads

A local server that bridges Claude's web interface (claude.ai) into an OpenAI-compatible REST API. Send requests to http://127.0.0.1:3000/v1/chat/completions and get responses back using your existing browser login session.

Works with OpenClaw, custom scripts, IDE plugins, and any tool that speaks the OpenAI chat completions format.

🦞 OpenClaw Integration

Perfect for OpenClaw users! Use Claude's intelligence with all your OpenClaw skills and automations - no API costs required.

Quick Setup for OpenClaw

# In your OpenClaw config:
llm:
  provider: openai
  api_base: "http://127.0.0.1:3000/v1"
  api_key: "dummy-key-not-used"
  model: "claude-sonnet-4.5"
  streaming: true

📖 Complete OpenClaw Integration Guide

Works With Any OpenAI-Compatible Tool

OpenClaw - Personal AI assistant with 100+ skills
Continue.dev - AI code assistant for VS Code
Custom scripts - Python, JavaScript, shell scripts
IDE plugins - Cursor and other OpenAI-compatible tools

Using OpenAdapter in your project? Open a PR to add it here!

How It Works

Your App / Client                  OpenAdapter                    claude.ai
─────────────────       ───────────────────────────────       ───────────────
                           Express server (:3000)
  POST /v1/chat/    ──>   Receives OpenAI-format req    ──>   Types prompt
  completions              Manages Playwright browser          into chat UI
                           Polls DOM for response
  OpenAI-format     <──   Extracts & converts HTML      <──   Claude generates
  JSON / SSE stream        to Markdown                         response

Launches Chromium with a persistent profile (session stays logged in across restarts)
Receives OpenAI-format chat completion requests over HTTP
Extracts the user prompt and any file attachments (images, documents)
Types the prompt into Claude's web UI and submits it
Polls the DOM for the response, streaming chunks via SSE if requested
Converts Claude's HTML response to clean Markdown
Returns an OpenAI-compatible response (or SSE stream)

Requirements

Node.js v18+
npm
A graphical desktop (the browser runs headful — no headless mode due to Cloudflare)

Setup

macOS

git clone <repo-url> open-adapter
cd open-adapter
npm install
npx playwright install chromium

Playwright downloads its own bundled Chromium — you don't need Chrome installed.

Windows

git clone <repo-url> open-adapter
cd open-adapter
npm install
npx playwright install chromium

Use PowerShell or Windows Terminal. Command Prompt (cmd) works too but PowerShell is recommended.

Linux

git clone <repo-url> open-adapter
cd open-adapter
npm install
npx playwright install chromium
# Playwright may prompt you to install system dependencies:
npx playwright install-deps chromium

On Linux, Playwright needs certain system libraries (libnss3, libatk-bridge, etc.). The install-deps command installs them automatically (requires sudo).

First Run (Login)

On the first run, the browser opens to claude.ai. You must log in manually once:

node server.js

The browser opens to claude.ai
Log in with your credentials
Your session is saved in .browser-profile/ and persists across restarts
The server is now ready at http://127.0.0.1:3000

Usage

Start the server

npm start        # Runs unit tests first, then starts the server (recommended)
npm run dev      # Starts the server directly, skipping tests

The server listens on http://127.0.0.1:3000.

Send a request (non-streaming)

curl http://127.0.0.1:3000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [{"role": "user", "content": "What is 2+2?"}]
  }'

Send a request (streaming)

curl http://127.0.0.1:3000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [{"role": "user", "content": "Explain recursion"}],
    "stream": true
  }'

Send images or files (base64)

The adapter supports OpenAI's multimodal message format:

curl http://127.0.0.1:3000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [{
      "role": "user",
      "content": [
        {"type": "text", "text": "What is in this image?"},
        {"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}}
      ]
    }]
  }'

CLI adapter (standalone)

The original CLI tool is also available for one-off queries without running the server:

node adapter.js "What is 2+2?"

Output goes to stdout. Status messages go to stderr, so you can pipe cleanly:

node adapter.js "List 5 prime numbers" | grep -i prime

Project Structure

open-adapter/
├── server.js               # Express API server (OpenAI-compatible endpoint)
├── adapter.js              # Standalone CLI tool (single-prompt, exits after)
├── lib/
│   ├── sessionManager.js   # Browser lifecycle & multi-tier session recovery
│   ├── extractPayload.js   # OpenAI message parser & file attachment handler
│   ├── htmlToMd.js         # HTML-to-Markdown converter (runs in-browser)
│   └── rateLimiter.js      # Detects Claude rate limits from DOM & response text
├── tests/
│   ├── unit/               # Unit tests (extractPayload, htmlToMd, rateLimiter)
│   └── integration/        # Integration tests (HTTP endpoint validation)
├── .browser-profile/       # Persistent Chromium session (created on first run)
├── temp_uploads/           # Temporary directory for file attachments
├── logs.txt                # Request/response logs (created at runtime)
├── package.json
└── README.md

Architecture

server.js — API Server

The main entry point. An Express server that:

Accepts POST /v1/chat/completions in OpenAI chat format
Supports both regular JSON responses and SSE streaming ("stream": true)
Handles multimodal content: text, base64 images (image_url), and file attachments (file_url)
Deduplicates system context across requests (hashes system messages, only re-uploads when changed)
Converts large prompts (>15k chars) to file attachments automatically
Logs all requests and responses to logs.txt
Returns OpenAI-shaped responses with estimated token counts

lib/sessionManager.js — Browser Lifecycle

Manages the Playwright browser with multi-tier recovery:

Level	Strategy	Description
L0	`isPageAlive()`	Quick JS eval liveness probe
L1	`reloadPage()`	Reload the current page
L2	`newChat()`	Navigate to `claude.ai/new` (fresh conversation)
L3	`restartBrowser()`	Close browser context + relaunch Playwright
L4	Fatal	Return null, server responds with 503

Sessions auto-timeout after 1 hour of inactivity, starting a fresh conversation.

lib/htmlToMd.js — HTML to Markdown

A self-contained DOM-to-Markdown converter that runs inside the browser via page.evaluate(). Handles headings, bold/italic, code blocks (with language detection), tables, lists, checkboxes, links, images, blockquotes, and horizontal rules.

lib/rateLimiter.js — Rate Limit Detection

Detects Claude's rate limiting by:

Scanning the DOM for error/alert elements
Pattern-matching the response text against known rate-limit phrases
Parsing retry-after durations from the message

Returns OpenAI-format 429 responses with Retry-After headers.

adapter.js — CLI Tool

The original proof-of-concept. A standalone script that launches its own browser, sends a single prompt, captures the response, and exits. Useful for quick scripting without running the server.

Configuration

Server (server.js)

Variable	Default	Description
`PORT`	`3000`	Server listen port
`MAX_TIMEOUT_MS`	`180000` (3 min)	Hard timeout waiting for Claude's response
`STABLE_INTERVAL_MS`	`30000` (30 sec)	Content-unchanged = done threshold
`POLL_MS`	`500`	DOM polling interval
`SESSION_TIMEOUT_MS`	`3600000` (1 hr)	Inactivity before starting a new conversation

CLI (adapter.js)

Variable	Default	Description
`CLAUDE_URL`	`https://claude.ai/new`	Starting URL
`USER_DATA_DIR`	`.browser-profile/`	Persistent browser session directory
`MAX_TIMEOUT_MS`	`120000` (2 min)	Hard timeout for response
`STABLE_INTERVAL_MS`	`3000` (3 sec)	Content-unchanged = done threshold
`POLL_MS`	`500`	Polling interval

Selectors

Both the server and CLI use fallback selector chains to find UI elements. If Claude updates their UI, edit SELECTOR_CHAINS:

promptInput:    div[contenteditable="true"] → div[role="textbox"] → ...
sendButton:     button[aria-label*="Send"] → button[data-testid="send-button"]
stopButton:     button[aria-label*="Stop"] → button[data-testid="stop-button"]
responseBlocks: div[data-testid*="message"] → div.font-claude-response → ...
fileInput:      input[type="file"]

To inspect current selectors, open claude.ai in Chrome DevTools (F12) and inspect the elements.

Testing

npm test              # Run all tests (unit + integration)
npm run test:unit     # Unit tests only (no server needed)
npm run test:integration  # Integration tests (requires running server)

Unit tests cover the core modules (extractPayload, htmlToMd, rateLimiter) and run automatically before the server starts when using npm start.

Integration tests validate the HTTP endpoint (request validation, CORS, response shape) against a live server.

Limitations

Headful only — a visible browser window is required (Cloudflare blocks headless)
Single request at a time — concurrent requests return 429 (busy)
No login automation — you log in manually once; session persists in .browser-profile/
Selectors may break — Claude UI updates can change the DOM structure
No conversation memory — each server session timeout starts a fresh chat
Token counts are estimates — calculated from character length, not actual tokenization

Troubleshooting

Problem	Fix
"Prompt input element not found"	Claude's UI changed. Inspect the page and update `SELECTOR_CHAINS`
Cloudflare challenge page	Must run headful (default). Don't set `headless: true`
Login not persisting	Ensure `.browser-profile/` exists and isn't being deleted
Timeout with no response	Increase `MAX_TIMEOUT_MS` or check if Claude is down
Browser doesn't open	Run `npx playwright install chromium`
503 session recovery failed	All recovery tiers failed. Restart `node server.js`
429 rate limit	Claude's free/pro message limit hit. Wait for the retry-after period
Linux: missing shared libraries	Run `npx playwright install-deps chromium` to install system dependencies
Linux: no display / `DISPLAY not set`	You need a graphical desktop. For headless servers, use Xvfb: `xvfb-run node server.js`
Windows: `npx` not recognized	Ensure Node.js is in your PATH. Reinstall Node.js using the official installer and check "Add to PATH"
macOS: Chromium blocked by Gatekeeper	Go to System Settings > Privacy & Security and click "Allow Anyway" for the Chromium binary

Contributing

We welcome contributions! OpenAdapter is built with the OpenClaw community in mind. 🦞

Good First Issues

Check out issues labeled good first issue - these are perfect for newcomers!

Priority Features

🎯 Tool Calling Support (v1.2) - #1 requested feature for OpenClaw integration
Docker support
Better streaming (MutationObserver)
Configuration file

See CONTRIBUTING.md for the full guide and roadmap for planned features.

Community

💬 Discussions - Ask questions, share workflows
🐛 Issues - Bug reports and feature requests
🦞 OpenClaw Integration - OpenClaw-specific help

License

ISC

Name		Name	Last commit message	Last commit date
Latest commit History 7 Commits
.github		.github
lib		lib
tests		tests
.gitignore		.gitignore
CLAUDE.md		CLAUDE.md
CODING_STANDARDS.md		CODING_STANDARDS.md
CONTRIBUTING.md		CONTRIBUTING.md
GITHUB_WEB_UI_TASKS.md		GITHUB_WEB_UI_TASKS.md
LICENSE		LICENSE
README.md		README.md
adapter.js		adapter.js
package-lock.json		package-lock.json
package.json		package.json
server.js		server.js

Folders and files

Latest commit

History

Repository files navigation

OpenAdapter - Claude.ai to OpenAI API Bridge for OpenClaw

🦞 OpenClaw Integration

Quick Setup for OpenClaw

Works With Any OpenAI-Compatible Tool

How It Works

Requirements

Setup

macOS

Windows

Linux

First Run (Login)

Usage

Start the server

Send a request (non-streaming)

Send a request (streaming)

Send images or files (base64)

CLI adapter (standalone)

Project Structure

Architecture

server.js — API Server

lib/sessionManager.js — Browser Lifecycle

lib/htmlToMd.js — HTML to Markdown

lib/rateLimiter.js — Rate Limit Detection

adapter.js — CLI Tool

Configuration

Server (server.js)

CLI (adapter.js)

Selectors

Testing

Limitations

Troubleshooting

Contributing

Good First Issues

Priority Features

Community

License

About

Topics

Resources

License

Contributing

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages