Ask
Ask in your assistant
A question or a request, in plain words.
- Client
- Claude
- Workspace
- Chosen at consent
- Request
- Scale the best ad set
MCP server
Your Meta ad work, inside your AI assistant. The Adrails MCP server lets Claude, ChatGPT, Claude Code, Codex and other remote MCP clients work on your Meta ad accounts. Sign in with OAuth, choose one workspace and its permissions, then ask. Reads return your figures, and launches, pauses and budget moves come back as exact changes you apply in Adrails.
Product view
Your assistant reads the same workspace you open in Adrails. Check a figure in Campaigns or Analysis, and open the change it prepared before it reaches Meta.

Product workflow
Your assistant reads through the server and prepares one exact change. You apply it in Adrails, and only then does it reach Meta.
Ask
A question or a request, in plain words.
Read
Read tools return your figures and change nothing.
Proposal
A write tool returns a proposal and its link, nothing more.
Apply
You review the exact change, signed in as yourself.
Illustrative workspace data. Values show the product flow and are not customer results.
MCP tools
Each name is the tool your client lists. A read changes nothing. A proposal returns a link to the change, applied from Adrails.
01
Try this prompt: List the Meta accounts in this workspace. For the account I choose, compare spend and conversions in the last seven complete days with the previous seven. State the dates, currency and time zone. Make no changes.
The assistant can use accounts_list to identify the account and performance_period for the two exact date ranges. Check that the result names the intended account and conversion definition before interpreting the trend. These tools require adrails:read; preparing advertising changes is a separate permission.
This is a reproducible starting check against the registered tool contract, not a claimed result from a customer account. If a figure is missing, confirm account access and the reporting period instead of treating it as zero. Use the setup documentation for connection steps and the comparison guide for choosing between servers.
02
Ask for spend, results, revenue, ROAS, CPA, CTR, CPM, reach and frequency over an exact period, at the account, campaign, ad set or ad level, and compare it with another period. The server reads Meta for that ad account and returns a table your assistant reasons over.
Other tools go past raw figures: they review campaigns against their actual objective, say whether each budget should be kept, raised, cut or investigated, and group ads by the image or video they use to find winners and fatigue.
03
A new Meta launch starts with launch_brief. It takes only what you said or chose and returns what a clean launch still needs, as questions with options: the ad account, objective, pixel and event, countries, budget, dates, Page and copy. Once nothing is missing, campaigns_launch prepares the campaign, ad sets and ads.
When you apply it, Adrails sends the images or Google Drive files to the ad account and builds everything paused. Switching the ads on is a second proposal, and it lists every budget that starts spending.
04
A pause, a resume, a budget change of at most 30%, a duplicate or a split test is prepared as one exact change. The result carries a link to the Review advertising change page in Adrails, where you apply it or decline it.
When it is applied, Adrails reads Meta again and applies the change once, or blocks it if the entity changed in the meantime. A proposal expires after 24 hours, and an applied change can be undone while Meta still matches what Adrails wrote.
05
Add https://adrails.ai/api/mcp to your client. It registers itself and opens Adrails in your browser, where the consent screen asks for the workspace and the permissions. There is no client ID, secret or API key to copy.
Read access is required and covers every ad account in that workspace. Preparing advertising changes is a separate permission that only an owner or admin can grant. Revoke a client at any time under Your connections on the MCP page.
A server that speaks the Model Context Protocol, so an AI assistant such as Claude or ChatGPT can call tools that read and act on Meta ad accounts. Adrails runs one at https://adrails.ai/api/mcp, and your assistant signs in to it with your Adrails account.
Claude on claude.ai and in the desktop app, Claude Code, ChatGPT with Developer mode and Codex are documented step by step. Any client that supports remote MCP over Streamable HTTP, with OAuth and dynamic client registration, connects the same way.
No. No tool creates, edits, pauses, switches on or rebudgets a campaign, ad set or ad by itself. It prepares the change, and only an owner or admin signed in to Adrails can apply it. The permission prompt in your AI client does not apply anything.
No. Meta is connected once in Adrails. The assistant signs in to Adrails with OAuth, and tool results never contain provider credentials.
No. MCP tool calls cost 0 credits, reads and changes alike: your assistant does the thinking. Applying a proposal needs a paid plan. Your assistant's subscription is separate.
Yes. A connection covers every ad account in its workspace, and a call can narrow to some of them. To work in another workspace, connect again and choose it.