--- title: 'AI agents for Meta ads | Plainrouter' description: 'Connect Claude Code, ChatGPT or Codex to Meta ads through Plainrouter. Start read-only, inspect evidence and approve supported changes in Ask mode.' canonical: 'https://plainrouter.com/agents' last_updated: '2026-10-08' --- # Your AI agent. Your Meta ads. Your approval. Claude Code, ChatGPT and Codex can review Meta ads through Plainrouter and propose supported changes. Sign up in Plainrouter, connect Meta and start with Read access. [Start for free](/register) · [Try the sandbox](/docs/mcp/setup#start-in-test-mode) ## Choose your agent ### Claude Code **Workspace key in your project configuration.** Use Claude Code’s terminal to review Meta account evidence alongside implementation work; start with Read access and approve proposals in Ask mode. [Use Claude Code for Meta ads](/agents/claude/meta). ### ChatGPT **OAuth sign-in in ChatGPT on the web.** Choose account and access during sign-in; developer mode requires an eligible paid plan, and workspace administrators may control availability. [Use ChatGPT for Meta ads](/agents/chatgpt/meta). ### Codex **Workspace key from your environment variable.** Review Meta ads in your existing Codex workflow, with a remote server configured in TOML and your credential kept outside configuration. [Use Codex for Meta ads](/agents/codex/meta). ## Other supported clients Setup guides also cover [Cursor](/docs/mcp/clients/cursor), [Gemini CLI](/docs/mcp/clients/gemini), [Grok](/docs/mcp/clients/grok), [Meta Muse](/docs/mcp/clients/muse), [Hermes](/docs/mcp/clients/hermes), [OpenClaw](/docs/mcp/clients/openclaw), [n8n](/docs/mcp/clients/n8n) and [Make](/docs/mcp/clients/make). These clients have documentation, but no dedicated Agents page yet. Follow the guide for the actual client: a CLI, web connector and workflow module have different setup steps. ## What every agent gets Start with reads of account context, retained performance, creative inventory and stored conversion-signal health. With Write access, supported proposals include pauses, resumes, daily-budget changes and creative tests. New ad copies start paused; a later governed resume can activate them. Ask mode waits for a person to approve each change. Full mode automatically queues changes allowed by policy. Both enforce policy, keep decision receipts and offer a kill switch to stop further changes; stopping does not undo earlier work. See [what a Meta ads AI agent can do](/solutions/meta-ads-agent). The account-free sandbox has four synthetic tools, stores nothing and never contacts Meta. It is a way to inspect read results, not to practice executing changes. ## How should you choose? Choose the client you already use and can configure safely. Claude Code and Codex use environment-held workspace keys; ChatGPT uses Plainrouter OAuth sign-in. Terminal configuration suits a developer workflow, while ChatGPT offers a web conversation with plan and administrator requirements. Client confirmation prompts and Plainrouter approval are separate controls, whichever client you choose. ## Frequently asked questions ### Do I need a new chatbot? No. Plainrouter connects your existing MCP client to supported Meta account workflows. You keep using that client; Plainrouter provides account access, policy and execution controls. ### Can I try it without connecting Meta? Yes. Connect the sandbox with no authentication and look for `sandbox: true` in a tool response. Its data is fictional, and it offers no production write tools. ### What does the connection cost? MCP is included on every Plainrouter plan, and MCP calls are not metered. Your AI client’s charges are separate; see [Plainrouter pricing](/pricing). ## Related guides - [What your Meta ads agent can do]() - [Hosted Meta Ads MCP overview]() - [AI agent safety checklist]() - [Connection reference]() - [Understand Marketing API access and limits]() - [Set up Meta’s official Ads MCP]() - [What the official server does not cover]()