The 7 Platforms
MultAI automates the web UIs of 7 AI platforms using your existing browser sessions — no API keys required for any of them:
| Identifier | Platform | URL | DEEP mode |
|---|---|---|---|
| claude | Claude.ai | claude.ai | — |
| chatgpt | ChatGPT | chatgpt.com | ✓ Deep Research |
| gemini | Google Gemini | gemini.google.com | ✓ Deep Research |
| grok | Grok (xAI) | grok.com | — |
| deepseek | DeepSeek | chat.deepseek.com | ✓ DeepThink |
| perplexity | Perplexity | perplexity.ai | — |
| copilot | Microsoft Copilot | copilot.microsoft.com | — |
Each identifier is used with the --platforms flag and is case-insensitive. Platforms without DEEP mode support receive the same prompt in REGULAR mode when DEEP is selected.
How CDP Browser Automation Works
The Chrome DevTools Protocol (CDP) is a low-level debugging interface built into Chrome that lets external processes inspect, control, and communicate with browser tabs. MultAI uses Playwright — a browser automation library — to connect to your existing Chrome instance on 127.0.0.1:9222.
When you run MultAI, the orchestrator:
- Connects to Chrome via CDP over the loopback interface
- Opens or reuses a tab for each platform
- Uses your existing browser session (cookies, login state) — no re-authentication needed
- Injects your prompt into the platform's input field and submits it
- Waits for the response to finish streaming, then extracts the text
- Saves the response to
output/and marks the platform complete
127.0.0.1 (not 0.0.0.0), so it's only accessible from your own machine. No remote access is possible.DEEP vs REGULAR Mode
MultAI supports two prompt modes, selected with the --mode flag:
REGULAR mode (default)
Sends your prompt as a standard chat message on all 7 platforms. Responses arrive in 30 seconds to 3 minutes per platform. Best for quick questions, comparisons, and tasks that don't require deep analysis.
DEEP mode
On platforms that support it, activates the extended research or extended thinking feature before submitting:
- ChatGPT — enables Deep Research mode (uses web search, takes 5–20 minutes)
- Gemini — enables Deep Research mode (uses Google Search, takes 5–15 minutes)
- DeepSeek — enables DeepThink (extended chain-of-thought reasoning)
Platforms without DEEP support (Claude, Grok, Perplexity, Copilot) receive the same prompt in their standard mode.
Rate Limiting
MultAI tracks per-platform usage and enforces cooldown periods to avoid hammering platform limits. Rate limit state is persisted at ~/.chrome-playwright/rate-limit-state.json and is updated after every run.
When a platform hits its rate limit during a run:
- The platform is gracefully skipped for that run
- The state file records the timestamp and cooldown period
- On subsequent runs, MultAI checks the state file and skips the platform until the cooldown expires
- Output for skipped platforms notes the rate-limit reason
To reset all rate limits manually, delete the state file:
Agent Fallback
If Playwright automation fails for a platform — for example, a UI change breaks the element selectors — MultAI can fall back to a browser-use agent that uses an AI model to navigate the interface in a more flexible, vision-based way.
Agent fallback requires one of:
ANTHROPIC_API_KEYin.env— uses Claude to drive browser-useGOOGLE_API_KEYin.env— uses Gemini to drive browser-use
Agent fallback is slower than Playwright (typically 2–5x) but significantly more resilient to UI changes. All fallback events are logged to agent-fallback-log.json in the output directory.
--with-fallback at runtime, or by running bash setup.sh --with-fallback to install the browser-use dependency.Parallel vs Sequential Execution
MultAI runs differently depending on how it's invoked:
Claude Code tab — parallel (asyncio)
The Python orchestrator uses asyncio to run all platforms concurrently. All 7 platforms start at the same time and complete independently. A full REGULAR-mode run typically finishes in 2–5 minutes regardless of platform count.
Cowork tab — sequential (Claude-in-Chrome MCP)
When invoked from the Claude Cowork tab, MultAI uses Claude-in-Chrome MCP browser tools instead of Playwright. Platforms run one at a time in sequence. No Python or Playwright is needed — just the Claude-in-Chrome browser extension. Sequential mode is slower but requires no local setup beyond the extension.
Output Files
MultAI produces two categories of output for each run:
Per-platform files (output/)
Each platform's response is saved as a separate markdown file:
Each file wraps the response in <untrusted_platform_response> XML tags, signalling to Claude that the content is external and should be treated with appropriate skepticism.
Collated archive (archive/)
After all platforms complete, collate_responses.py merges all per-platform files into a single archive:
The archive is the primary artifact — it contains all 7 responses in a single scrollable document, ready for review or synthesis with /consolidator.