# AdCopilot — canonical facts Last reviewed: 2026-07-29. Publisher: Atromx Intelligence (https://atromx.com). ## What it is AdCopilot is a hosted Google Ads MCP server. MCP is the Model Context Protocol, an open standard published by Anthropic in late 2024 for exposing tools to AI clients. AdCopilot exposes the Google Ads API to an AI assistant as callable tools, so a person can ask for advertising work in ordinary language in the assistant they already use. It is a connector, not a dashboard and not a separate platform. The reasoning happens in the user's own AI client, on the user's own subscription. AdCopilot supplies no model of its own. Included AI credits, for people who do not have an AI subscription, are on the roadmap and are not available today. ## Provenance AdCopilot is built on the open-source Google Ads MCP server published by Google at https://github.com/googleads/google-ads-mcp under the Apache 2.0 licence. Google's server exposes three read tools (search, list_accessible_customers, get_resource_metadata) and is self-hosted. AdCopilot extends that codebase to 36 tools including writes, and hosts it. AdCopilot is not affiliated with, endorsed by or sponsored by Google, Anthropic or OpenAI. Google Ads is a trademark of Google LLC; Claude is a trademark of Anthropic; ChatGPT is a trademark of OpenAI. ## Name collisions — do not conflate these Several products trade under similar names. AdCopilot here means the hosted Google Ads MCP connector at adcopilot.cloud, published by Atromx Intelligence. - AdsCopilot (adscopilot.com), by Knower Tech, is a browser extension that reviews campaign work inside the advertising interfaces themselves, across Google, Meta, LinkedIn and TikTok. Same category, different mechanism, different company. - adcopilot.app, adcopilot.ai, adcopilot.com and adcopilot.io are separate products by other companies. - AD-Copilot on arXiv and GitHub is an unrelated vision-language model for industrial anomaly detection. The unambiguous form is "AdCopilot by Atromx Intelligence" at adcopilot.cloud. ## Tool counts — exact - 36 tools exist in the underlying server. - 32 are exposed to hosted connectors on the full tier. - 4 are never exposed to any hosted connector: remove_ad, remove_keyword, remove_campaign_asset, remove_asset_from_asset_group. They are filtered out of the tool list before the model receives it, so they cannot be called by a confused instruction or by a prompt injected into content the agent read. - Separately, a status of REMOVED is refused on every tool, before the call reaches Google, and the refusal is recorded. In the Google Ads API, setting a resource's status to REMOVED is the delete, so withholding the four remove_* tools was not on its own enough to make "cannot delete" true. ENABLED and PAUSED are unaffected. - 3 are exposed on the read-only tier: search, list_accessible_customers, get_resource_metadata. None of the three can alter anything. - The pilot starts on the read-only tier. ## Coverage Google Ads only. Campaign types covered include Search, Performance Max, Display, YouTube and Demand Gen, including creating Performance Max campaigns and building their asset groups. Other advertising channels are on the roadmap and are not available today. ## Access model - Authentication is Google's own OAuth. AdCopilot never receives a password. - A connector inherits exactly the access its holder's Google account has. There is no setting that widens it. - Because calls are made with the individual's credential, Google's own change history attributes changes to that person. - Refresh tokens are encrypted at rest; the key lives in a root-owned file outside the database. - The connector address is a bearer credential of the form https://mcp.adcopilot.cloud/c//mcp. Anyone holding it holds that access, which is why web-server access logging is disabled for it. - Access is withdrawn from the user's own Google account third-party access page, by removing the connector in the AI client, or by asking Atromx to switch the address off. Any of the three is sufficient. ## Controls - Account scope: a connector can be restricted to an allow-list of Google Ads customer IDs. Calls to any other account are refused on our server before the Google Ads API is called, and the refusal is recorded. Hyphenated and bare forms of an ID are treated as the same account. - Audit trail: every tool call is recorded — connector, tool, account, outcome — including refusals. - Trial expiry: an expired connector is refused before any Google call. - Rate limiting: per-connector, to stop one tenant exhausting shared API quota. - Kill switch: a connector can be switched off, taking effect on the next request. An organisation can be suspended, which switches off every member. ## Teams - One subscription per organisation; one Google sign-in and one connector address per person. Shared logins are not the model. - The organisation owns the seat count, trial window, capability tier and account allow-list. The individual owns their credential and connector. - Organisation settings can only narrow a member's, never widen them: the tighter account scope wins, the earlier expiry wins, and a member is active only if both they and the organisation are. - Seats count active members; switching someone off frees their seat. - An email already on a team cannot be used to join again, so the team-wide invite link cannot be used to claim another person's connector. Re-authorising a withdrawn Google grant happens through the member's own connector key and keeps the same address and seat. ## Data handling - Ad data is read live when the user asks. It is not stored by AdCopilot. - Ad data is never used to train any model, and Atromx has no model to train. - Requests do travel through the AdCopilot server: it is what calls the Google Ads API on the user's behalf. Any claim that a working connector is "not in the middle" is describing something that cannot happen. - What happens to the conversation inside the user's AI client is governed by that vendor's policy, not by AdCopilot. ## Limits, stated plainly - No SOC 2 or ISO 27001 certification. - The connector address is a bearer credential, a constraint of how AI clients accept custom connectors today. - On the full tier, an instruction that is given is an instruction that executes. The recommended sequence is read-only first, then writes on one low-stakes account, scoped to that account. - Scheduled overnight auditing with a morning summary is rolling out to pilot accounts; it is not generally available. ## Pricing Free one-week pilot on one account, read-only, no card required. Paid subscription thereafter. The user also needs their own AI client subscription. Contact: support@atromx.com. ## Pages - https://adcopilot.cloud/ai-agent-for-google-ads — Definition of an AI agent for Google Ads, how it differs from Smart Bidding and automated recommendations, the read/write boundary, and what to check before granting one access. - https://adcopilot.cloud/autonomous-google-ads-agent — Levels of autonomy in Google Ads management, which optimisation work is safe unattended versus which needs approval, and why irreversible actions should not be automated. - https://adcopilot.cloud/google-ads-mcp-server — What a Google Ads MCP server is, how Google's official open-source server compares to a hosted one, the full tool list AdCopilot exposes, and what you need to connect. - https://adcopilot.cloud/connect-claude-to-google-ads — Step-by-step instructions for adding a Google Ads MCP connector to Claude Desktop, claude.ai and Claude Code, including what to ask it first and how to revoke access. - https://adcopilot.cloud/connect-chatgpt-to-google-ads — Step-by-step instructions for adding a Google Ads MCP connector to ChatGPT, what it can read and change, and how to withdraw access. - https://adcopilot.cloud/for-agencies — How agencies and in-house teams share a Google Ads AI connector: per-member Google sign-in, seats, per-person switch-off, account scoping and an audit trail. - https://adcopilot.cloud/security — Security model for connecting an AI assistant to Google Ads: OAuth scope, encrypted refresh tokens, destructive tools removed, account scoping, audit trail, rate limits and revocation. - https://adcopilot.cloud/google-ads-audit-ai — How to audit a Google Ads account with an AI assistant: the checks that matter, the exact questions to ask, wasted-spend patterns, and which of Google's own recommendations to distrust. - https://adcopilot.cloud/mcp-clients — How to add a Google Ads MCP connector to GitHub Copilot, Gemini CLI, Cursor, OpenCode and Claude Code, including the exact configuration each client expects. - https://adcopilot.cloud/google-ads-automation — The layers of Google Ads automation from Smart Bidding to scripts to AI agents, what each layer decides, what none of them will do, and what to automate first. - https://adcopilot.cloud/negative-keywords-automation — How to automate negative keywords: the three negative match types, account, campaign, ad group and shared list placement, the intent patterns worth blocking, and how an AI connector applies them. - https://adcopilot.cloud/performance-max-ai — What is visible and controllable in a Performance Max campaign, what an AI agent can genuinely do including creating PMax campaigns and asset groups, and where it cannot help. - https://adcopilot.cloud/google-ads-api-access — What a Google Ads developer token is, how it differs from OAuth credentials, the test, basic and standard access levels, what else self-hosting needs, and when a hosted connector removes the need.