These answers cover Claude and the MCP server.

You make the decisions that commit money. AdBuyMCP balances your budget across the four channels, and nothing is bought until you've approved the plan, the ads and the start date. On this website, Claude, Anthropic's AI model, drafts who buys from you from your public homepage, and you correct the draft. In the product, an AI agent can prepare and propose changes, and a person approves anything that commits money.

Yes, with a workspace approved for MCP access and a compatible client whose own OAuth client ID and exact callback have been registered. AdBuyMCP runs the hosted server; you sign in through Supabase, select your workspace and consent. Not every Claude product has been registered or accepted yet, so confirm the client you use before connecting. The connection follows your workspace's enabled capabilities and grants no spend authority, and nothing is bought until the plan, the ads and the start date are confirmed with you. There is no public connector or registry listing.

It's a feature. AdBuyMCP buys awareness advertising, and the MCP server is one way to work on your plan: from Claude, or another approved client, instead of the browser. Both go through the same routes and approval records, so an agent gets exactly the rules a person gets.

No. The block is built into how the server authorises a call. Consequential calls fail closed unless the incoming request matches an approval a human granted for those exact bytes, spend runs against a mandate with an explicit money ceiling and expiry, and enabling spend on a live line is a second action granted separately from the approval. Authority is granted, withdrawn and closed by people; approvals are resolved by people.

The run stops. The exception waits for a person to read the response, say what happened and resume. That is the specific failure this design exists to prevent: an agent that cannot read a supplier's response and tries again has just created a second campaign nobody knows about.

The modern 2026-07-28 protocol, with server/discover negotiation and per-request metadata. The hosted service is reached over streamable HTTP, and the server also speaks stdio. The pre-2026 initialize handshake also works over stdio and stateless HTTP, and tests pin that compatibility path, including a 2025-11-25 HTTP round trip and cancellation behaviour in both eras.

Yes. The hosted endpoint uses standard streamable HTTP and Supabase OAuth. A compatible client needs protected-resource discovery, PKCE S256, resource audiences and an approved pre-registered client identity with exact callbacks. The protocol is shared; acceptance of one client does not prove another can connect. No client receives authority to approve its own spend.