Meta Ads
TikTok Adsplanned
Google Ads
Shopify
WooCommerceplanned
GA4
WordPress
HubSpotplanned
Mailchimp01
Same identity and permissions
An assistant acts as a person with a role, a client scope and rate limits, and cannot do more than that role can.
An MCP server with the same rules as the web app
Assistants read metrics, draft batches and preview changes through MCP, and can run an approved change only with the same permissions, limits and log as a person.
Floating cards and figures are examples on sample data, not a customer account.
1The proof
Every number comes with its receipt.
Sample session, fictional brand. This is how MCP for AI assistants shows a figure: the value, where it came from, the formula behind it and when it was last updated.
query_metricsMER, 30 daysSource and formula3.42 with formula and timeUpdatedsee source
draft_ad_batch3 ads, pausedSource and formulaDraft onlyUpdatedsee source
execute_approved_changeWithout approvalSource and formulaRefused: approval requiredUpdatedsee source
Rule: every number shows its source, formula and update time. A missing source reads "not measured", never 0.
2See it move
Channels light up one by one.
Each connector starts read-only. Built ones light up here, planned ones stay dim until they exist.
3Part of Operate
MCP for AI assistants: what it does.
One of four areas of the product, sharing the same client record, roles and log.
02
Read, then draft, then execute
Read tools come first. Draft and preview follow. Execute exists only for changes that have been approved.
03
No raw HTTP tool
There is no tool that exposes your tokens or sends arbitrary requests.
Works with
1 platform, read-only first
Connections start read-only. Writing is switched on per account, only after you grant the rights.
At a glance
Stated up front
- things it does
- 3
- steps start to finish
- 4
- platform connections
- 1
- limits stated
- 2
4The data path
How the data moves. Source to screen.
Lines draw as you scroll. Platforms on the left are read-only first, MCP for AI assistants sits on the one client record, and each output keeps its receipt.
RankfinityRead-only firstROASfinityMCP for AI assistants
Same identity and permissions
Read, then draft, then execute
No raw HTTP tool
Sources are read first, the numbers are stored once per client, and every output shows its source, formula and update time.
How the work moves in Operate
The same four moves apply to every module in this area.
- 1 CaptureFrom the work itselfTasks start from comments, results, approvals and requests.
- 2 AssignRoles and ownersOne owner per task, and clients only see what they should.
- 3 TrackStatus in one recordDue dates, SOP steps and approvals sit with the client record.
- 4 LogNothing is lostEvery change is written to a log that cannot be edited afterwards.
5The routine
How it works, step by step.
From first connection to the number a client sees.
- 1
Connect over HTTP
Add the server to your assistant.
Step 1Connect over HTTPexample - 2
Read
List clients, query metrics, inspect campaigns, search the library.
Step 2Readexample - 3
Draft and preview
Prepare an ad batch or a change and see the exact diff.
Step 3Draft and previewexample - 4
Execute when approved
An approved change runs once, is read back and is logged.
Step 4Execute when approvedexample
6In use
Built for businesses like these. Example scenes.
Photos are stock images that show the kind of business this is built for. They are examples, not customers.
Stock photos and example labels, not customers.
7In practice
Common mistakes, and a short checklist.
General practice for this kind of work, not a claim about any one tool.
Starting with write access
Connect read-only first and widen access for a reason you can write down.
One connection for every client
Scope each connection to the clients it needs so data does not cross.
No review of old connections
List connections each quarter and remove what nobody can explain.
A short checklist
Your ticks stay in this browser only.
8Honest limits
What it will not claim, and what it works with.
What it will not claim
- Documents, comments and web pages that an assistant reads are treated as data, never as instructions.
- The tool set grows with the connectors, so write tools exist only where a connector supports them.
Works with
Connections are read-only first and go live only after the platform access is approved.
Built: the connector exists in the product and goes live per account once access is granted. Planned: not built yet.
Used instead of
No single tool replaces this module; it connects work that is usually split across several.
9Related modules
Same room, next door.
Modules that share this client record, these roles and this log.
LaunchChange logEvery change with before and after, author, reason, time, API response and what happened next.Read more→OperateSEO with RankfinityRankings, local map scans and AI visibility from Rankfinity, attached to the right client and site.Read more→MeasureReportsClient, channel, campaign and product reports where every figure shows its source, formula and update time.Read more→
10Guides
Learn more. Short guides.
Short guides on the numbers and routines behind this module.
GuideA safety checklist before an AI assistant touches ad accountsAn assistant that can read your accounts is a convenience. One that can write to them is a decision. This checklist separates the two.6 min read→GuideAn approval workflow for changes to live ad accountsMost expensive mistakes in ad accounts are small edits applied to the wrong thing. A short, boring workflow prevents most of them.6 min read→
11FAQ
Questions about mcp for ai assistants.
Which assistants work?
Any MCP client, for example Claude or Codex, over the HTTP transport.
Can an assistant spend my budget?
Only through an approved change with the same limits as a person. Nothing runs without approval.
How does it relate to the Rankfinity MCP?
It can federate it, so SEO results are attached to the same client and site.
Run the next client review from one screen.
We open one agency at a time, with your own accounts connected read-only first.