By the end of this guide, Claude will be connected to AdStellar through the MCP server at https://app.adstellar.ai/api/mcp, so you can analyze, create, and launch Meta campaigns by chatting instead of bouncing between Ads Manager and dashboards. The setup takes a few minutes. The prompts are where the time savings come from, so this guide includes copy-paste examples for the full loop.
Before you start, you need an AdStellar account with your Meta ad account connected, plus either Claude on a plan that supports custom connectors or Claude Code installed on your machine.
Step 1: Prepare your AdStellar account and Meta connection
Claude can only act on what AdStellar can see, so the groundwork happens in AdStellar, not in Claude. Log in and confirm your Meta ad account shows as connected. If the connection has lapsed, reauthorize it now, because a stale token will surface later as confusing empty results.
Then work through this short checklist:
- Import existing campaigns. Pull in your current Meta campaigns so Claude can analyze real history from the very first prompt. Without history, your first analysis will have nothing to rank.
- Pick the right workspace. AdStellar supports unlimited workspaces, so agencies should open the workspace for the client they want to work on. Claude operates on the account tied to your session, and it is much easier to choose correctly now than to unwind a mistake later.
- Write your goal in one sentence. For example: "Lower CPA on my prospecting campaign without cutting spend." A single clear goal makes every later prompt sharper.
- Set a budget ceiling. Decide the monthly spend you are comfortable letting the workflow manage, and note a daily cap that fits. You will state this in your prompts later.
Keep the goal and budget in a note. You will paste them into prompts in Steps 5 through 7.
Step 2: Understand what MCP does and what you are connecting
MCP stands for Model Context Protocol, an open standard that lets an AI assistant call tools in an external app. Instead of you copying data into a chat, the assistant sends a structured request to the app and gets a structured answer back. In this setup, the app is AdStellar, and the tools cover analysis, creative generation, and campaign launch.
Everything in this guide uses one endpoint, with HTTP transport:
https://app.adstellar.ai/api/mcp
This clears up a common misconception. Claude is not guessing from screenshots or reading pasted CSV exports. When you ask about your campaigns, Claude calls AdStellar's tools, and AdStellar pulls real-time data from Meta. Claude never sees your Meta account directly. It works through what AdStellar exposes, which is why Step 1 matters so much.
That also means answers are only as good as your connected data. If a campaign is not imported, Claude cannot analyze it.
One caveat on Claude's side: connector menus, plan requirements, and the authorization flow change as Anthropic updates its products. As of November 2026, remote MCP servers are supported in Claude Code and through connector settings in Claude's apps, but check Claude's official documentation for current plan availability and menu names before you begin. The next two steps are two routes to the same result, so choose the one that matches how you use Claude.
Step 3: Add the AdStellar MCP server in Claude Code
If you work in a terminal, this is the fastest route. Run this exact command:
claude mcp add --transport http adstellar https://app.adstellar.ai/api/mcp
Each part does a specific job:
- claude mcp add registers a new MCP server with Claude Code.
- --transport http tells Claude Code to reach the server as a remote HTTP endpoint rather than launching a local process.
- adstellar is the name you give the server. It appears in lists and in the tool names Claude uses, so keep it short and recognizable.
- https://app.adstellar.ai/api/mcp is the AdStellar endpoint.
Verify the connection
Run claude mcp list in your terminal. You should see adstellar in the output. Then start a Claude Code session and type /mcp to view server status. If the server asks you to sign in or authorize, complete that prompt in the browser window it opens and approve access to your AdStellar account.
If it fails
Three causes account for most failed connections:
- A typo in the URL. Compare it character by character, including /api/mcp at the end.
- The wrong transport flag. Leaving out --transport http, or using a different transport, points Claude Code at the wrong connection method.
- An unfinished authorization. If /mcp shows the server as needing authentication, finish the sign-in. The server will not expose tools until you do.
If you need to start over, remove the server with claude mcp remove adstellar and rerun the add command.
Step 4: Add AdStellar as a connector in Claude (desktop or web)
This step is an alternative to Step 3, not an additional requirement. If you set up Claude Code, skip ahead to Step 5.
For Claude on desktop or web, you add AdStellar as a custom connector, which is Claude's name for a remote MCP server. Menu labels and plan availability change, so confirm them in Claude's documentation first. As of November 2026, the general flow looks like this:
- Open Claude's settings and find the connectors section.
- Choose the option to add a custom connector.
- Name it AdStellar and paste https://app.adstellar.ai/api/mcp as the server URL.
- Save, then follow the authorization prompt to sign in to AdStellar and grant access.
- Open a new conversation and enable the AdStellar connector from the tools or connectors control in the chat.
On some plans, adding connectors may be limited to account owners or admins. If you do not see the option, that is usually a plan or permission issue rather than a problem with the URL.
Confirm it works
In a conversation with the connector enabled, ask:
Which AdStellar tools can you use? List them and briefly describe what each one does.
Claude should return a list covering things like campaign analysis, creative generation, and launching. If it says it has no such tools, the connector is probably not enabled for that specific chat.
Step 5: Run your first analysis prompt
Start with analysis. It is read-only, so nothing can go wrong, and it proves the data flow before you let Claude build anything. Paste a prompt like this:
Analyze the last 30 days of my Meta campaigns. Which ads are fatiguing, and which creative elements correlate with my lowest CPA?
A good answer has four parts:
- A ranked list of winners and losers, not just a summary paragraph.
- Fatigue flags on specific ads, based on AdStellar's fatigue detection.
- Element-level scoring, so you can see whether headlines, hooks, images, or formats are driving results.
- Reasoning you can check. Pick two or three ads from the ranking and compare spend, CPA, and results against Meta Ads Manager for the same dates. They should line up.
The most common mistake is a vague prompt. "How are my ads doing?" gives Claude no metric and no window, so you get a generic answer. Always name the goal metric (ROAS, CPA) and the date range. Better still, include your one-sentence goal from Step 1:
My goal is lower CPA on prospecting. Using the last 30 days, rank my prospecting ads by CPA and flag any with rising frequency and falling click-through rate.
Drill down with a follow-up
Once you have the overview, narrow it:
For the top three ads by CPA, break results down by audience and placement. Where is the CPA gap largest, and what does that suggest?
Follow-ups like this are where the chat format beats a dashboard, because each answer shapes the next question.
Step 6: Create a campaign with a prompt
With a trustworthy analysis in hand, move to building. AdStellar's Agent can assemble a data-backed campaign from your past performance in under 60 seconds, so your prompt should point it at your proven material. Fill in the brackets with your own details:
Using my top-performing ads from the creative library, build a prospecting campaign for [product] with a $[X] daily budget and a target CPA of $[Y]. Do not launch it. Show me a draft summary.
Request new creative
If your winners are showing fatigue, ask for fresh assets in the same prompt or a follow-up:
- Generate three image ad concepts and two video ads for [product], following the angles that scored highest in the analysis.
- Create a UGC avatar ad with a talking-head presenter using the hook from my best-scoring ad.
- Build ads from this product page using URL to Ad Maker: [product URL].
Ask for the reasoning
AdStellar shows the reasoning behind each campaign, and Claude can relay it. Ask directly:
Explain why you chose these audiences, this budget split between ad sets, and these specific creatives. Which data points drove each choice?
If an explanation does not match what you saw in Step 5, push back and ask for an alternative.
Choose variations over a single ad
One ad tells you little. Ask for a bulk set instead:
Create 5 hooks, 4 headlines, and 2 formats for this campaign and combine them into a test set. Show me the full list before anything goes live.
AdStellar can launch hundreds of variations in minutes, so the limit is your review time, not production time. Read the draft before moving on.
Step 7: Launch, monitor, and iterate from the same chat
Launching is the step with real spend attached, so build in a confirmation. State your limits and ask Claude to wait for your approval:
Launch the approved campaign to Meta, starting paused, with a daily budget cap of $[X] and a total cap of $[Z] for the first week. Confirm the settings back to me and wait for my "go" before activating.
Starting paused lets you inspect the campaign inside Meta Ads Manager before a dollar is spent. Check the budget, audiences, placements, and ad copy there, then tell Claude to activate.
Monitor and replace
Once live, use ongoing prompts such as:
Check performance daily against my $[Y] target CPA. Flag fatigued creatives and suggest replacements from my ranked winners. Do not change budgets without asking me.
AdStellar monitors campaigns around the clock, so these check-ins reflect current Meta data rather than yesterday's export.
Safety habits worth keeping
- Put the budget cap in every launch or edit prompt, not just the first one.
- Read the draft summary before approving, every time.
- Spot-check the live campaign in Meta Ads Manager during the first few days.
- Require approval for budget changes until you trust the pattern.
Two optional upgrades help over time. Pair AdStellar with Cometly attribution for cleaner conversion data feeding your analysis, and save new winners back to the creative library so the next campaign starts from stronger material.
Where to go once the loop is running
First confirm the basics: the adstellar server shows as connected in claude mcp list or your connector settings, and your first analysis matches the numbers in Ads Manager. If both hold, you have a working link between Claude and your Meta data.
From there, expand gradually. Run analysis prompts for a week, move to drafts you review line by line, then graduate to launch prompts with caps in place. As trust builds, you can hand over more of the routine work, like fatigue checks and creative refreshes, and spend your own time on strategy and offers.
