AIrlock Join waitlist

Your AI asks. You approve. Nothing else leaves.

AIrlock sits between Claude, ChatGPT or Gemini and your Gmail, Drive and WhatsApp. The AI searches; you approve what it gets to see.

ClaudeChatGPTGeminibeta · via review link GmailGoogle DriveWhatsApplocal cacheLocal filesAny MCP server
Live demo

See what the AI sees. Before it sees it.

Claude searches your inbox. The results land on your screen, not in the model. Untick, redact, release.

New chat · AIrlock connected

When does Emma's flight land?

gmail_search via AIrlockwaiting for you to review r_8f2c
{"q": "from:emma flight after:2026/11/01"}
Review before releaseE2E encrypted · AI can't see this
    Exactly what the AI will receive
    
                

    Emma's flight NB 482 from New York lands at Heathrow T2 at 18:40 on Friday.

    1. Ask
    2. AI searches
    3. You review
    4. Redact
    5. Release
    6. AI answers
    7. scroll to step · fictional data
    The problem

    Because today it's all or nothing.

    Once your inbox is connected, the model can read anything it searches for: contracts, medical results, bank alerts, private conversations. One vague query and the whole thread lands in its context window.

    With AIrlock the AI still writes the queries (it's good at that). The difference is that you are the filter. Results go to your screen first, and only what you release goes back to the model.

    How it works

    It's an MCP proxy with a human (you!) in the loop.

    Your AI app gets exactly one custom connector: apl-local, running on your machine. It connects to your providers, re-exports their tools, and holds every result until you review it.

    1. Ask as usual. The model writes the query, e.g. gmail_search(q="from:emma flight").
    2. Review in the chat. apl-local runs it, keeps results in memory, and opens a panel inside the chat. You tick, trim, redact.
    3. The AI finishes the job with only what you released. Raw results are discarded once delivered or expired.