Running Meta ads from a chat window is now a practical workflow. Once Claude is connected to AdStellar through MCP, you can ask it to pull performance data, explain what is working, build a campaign from your best ads, and launch it, all in plain language. By the end of this guide, you will have that connection live and a set of copy-ready prompts for each stage. Before you start, you need an AdStellar account with your Meta ad account connected, plus a Claude plan or client that supports custom MCP connectors. Claude never touches Meta directly. AdStellar is the bridge, and Claude works through the tools it exposes.
Step 1: Prepare your AdStellar workspace and Meta connection
Claude can only act on what AdStellar can see, so fix the foundation first. Log in to AdStellar and confirm your Meta ad account shows as connected and that spend, results, and recent ads are loading. If the dashboard looks stale or empty, resolve that before touching Claude, because every answer it gives will inherit the gap.
Next, import your existing Meta campaigns. Past performance is the raw material for everything later: analysis of winners and losers, creative scoring, and campaigns built from proven elements. A fresh account with no history still works, but Claude will have little to learn from.
If you manage more than one business, pick the correct workspace. AdStellar supports unlimited workspaces, so keep each brand or client separate. Agencies should connect one client workspace at a time. That habit prevents the most expensive kind of mistake, where a prompt meant for one client's budget lands in another's.
Finally, check that your Claude setup allows custom connectors. As of November 2026, availability depends on your plan and client, and these rules change often, so confirm in Claude's current settings or help documentation. On team or organization plans, an admin may need to enable custom connectors before you can add one. Sorting that out now saves you a dead end in the next step.
- Meta ad account connected and showing current data
- Existing campaigns imported
- Correct workspace selected
- Custom connectors available on your Claude plan or client
Step 2: Add the AdStellar MCP server URL in Claude
MCP, the Model Context Protocol, is an open standard that lets an AI assistant like Claude call external tools and read data from a service. In this setup, AdStellar runs the MCP server, and Claude connects to it as a custom connector.
Open Claude's settings and look for the connectors or integrations area. Choose the option to add a custom connector. Menu labels and locations differ by client and shift over time, so as of November 2026 treat the names here as approximate and follow the labels on your screen.
When the form asks for details, enter these:
- Name: AdStellar
- URL:
https://app.adstellar.ai/api/mcp
Use the URL exactly as written, with no extra path segments, no trailing parameters, and no alternative address. The MCP endpoint is this one URL and nothing else. Copy and paste it rather than typing it, and check that no space or line break came along for the ride.
Common mistake: typos and substituted endpoints. If you paste the plain AdStellar login address, a dashboard link, or a URL from some other tool, the connector will either fail to save or save and expose nothing. A trailing slash or appended query string can cause the same trouble. If the connector will not add, compare your entry character by character against the URL above.
Save the connector. At this point Claude knows where the server is, but it cannot read your account until you authorize access in the next step.
Step 3: Authorize access and verify the connection
After saving, Claude should prompt you to connect or sign in. Complete the authorization with the AdStellar account that owns the workspace you prepared in Step 1. The exact wording and screens of the flow can vary, so follow what appears on screen. Read the permissions requested before you approve. You are granting an assistant the ability to act in your ad workspace, so confirm you are comfortable with that scope.
Then open a new chat and enable the AdStellar connector. In most clients this is a tools or connectors toggle near the message box. A connector that is installed but switched off in the current chat will not be used.
Run a harmless, read-only test:
List the tools you can use from AdStellar and show my connected ad account name.
A working connection returns a list of available tools and the name of the ad account you expect. Check that the account name matches the workspace you intended. If it shows the wrong one, stop and fix the workspace before going further.
If no tools appear
- Re-check the URL against
https://app.adstellar.ai/api/mcp, character by character. - Reauthorize the connector, signing in with the correct AdStellar account.
- Start a fresh chat, since a conversation opened before setup may not pick up new tools.
- Confirm the connector is toggled on for that chat.
- On a team plan, ask your admin whether custom connectors are permitted.
Step 4: Analyze your Meta account performance with prompts
Start with read-only analysis. It builds your trust in the connection and shows you how Claude reasons about your data before it can change anything.
Be specific about three things: the date range, your goal metric, and your attribution window. "How are my ads doing?" invites a vague answer. This is better:
Using my connected Meta account, show the top 5 and bottom 5 ads by ROAS over the last 30 days. My attribution window is 7-day click. Include spend, purchases, and CPA for each.
Then go after waste and decay:
Analyze the last 30 days. Which ad sets are spending without conversions, and which creatives show fatigue signals?
To get beyond account-level numbers, ask for element-level insight. AdStellar's creative analytics scores individual ad elements, and its AI Insights can surface patterns across them:
Break down my best-performing ads by hook, visual style, copy angle, and CTA. What do the winners share that the losers do not?
Always make Claude show its work. Add a line such as: Explain your reasoning and list the exact figures you pulled for each claim. Then spot-check several of those numbers against Meta Ads Manager. Attribution settings, date ranges, and time zones can make figures differ slightly, so verify that the gap is explainable. Do not act on any number Claude cannot source. If a figure has no citation, ask where it came from, or ask it to pull the data again.
Step 5: Create a data-backed campaign and creatives
With analysis done, move to building. Give Claude a complete brief so it does not fill gaps with guesses. Include:
- Campaign objective
- Product or landing page URL
- Daily or total budget
- Audience constraints and geography
- KPI target, such as a maximum CPA or minimum ROAS
A working prompt looks like this:
Build a conversion campaign for [product URL] using my best-performing past ads. Daily budget $X, target CPA $Y, targeting [country], excluding existing customers. Explain which past ads and audiences you chose and why.
Before generating anything new, ask Claude to check the ranked creative library for proven winners. Reusing an ad that already earned its place beats starting from zero:
Which creatives in my library should be reused for this campaign? Rank them by performance against my goal metric.
Where you need fresh creative, say what you want. AdStellar can produce AI image ads, video ads, and UGC avatar or talking-head ads. For a starting point, you can ask for an Ad Clone based on a reference ad, or use URL to Ad Maker to turn a product page into ad concepts. Specify format, placement, and angle in the prompt.
Before any bulk generation, ask for a variation plan:
Propose a variation plan first: hooks, headlines, and formats, with the number of ads each combination produces. Do not generate yet.
Editing a plan is quick. Cleaning up three hundred ads built on the wrong structure is not. Cut weak angles and adjust the mix now, then tell Claude to proceed.
Step 6: Review the plan, then launch on Meta
Nothing should go live until you have seen the whole picture. Ask Claude to lay it out:
Show the full campaign plan: structure, budget per ad set, audiences, placements, and the complete creative list. For each choice, give the reasoning.
Read it as you would a media buyer's proposal. Does the budget match what you said? Are the audiences and exclusions right? Are the placements ones you want? Are any creatives off-brand?
Then decide how to launch. For a new setup, a staged first launch is the safer path: have the campaign created paused or as a draft so you can inspect it in Meta before spending begins. Ask Claude which launch options AdStellar offers, since the available states and defaults can change, and confirm rather than assume that a paused option exists. Once you trust the workflow, launching live directly is reasonable for routine campaigns.
When you are ready, restate the key terms in the approval:
Launch this as planned, using bulk launch for all [N] variations. Daily budget $X, audience [description]. After launching, confirm exactly what was created.
Mistake to avoid: a bare "go ahead." An approval that does not repeat budget and audience leaves room for misreading, and you lose a clear record of what you authorized.
Afterward, open Meta Ads Manager and check that the campaign, budget, audiences, and ad count match what Claude reported. If anything differs, ask Claude to explain the discrepancy before you change anything.
Step 7: Monitor, optimize, and iterate with follow-up prompts
A launched campaign needs a review rhythm. Pick one, such as a check at day three and a fuller review weekly, and prompt Claude each time:
Compare the new campaign to my benchmark CPA of $Y after 3 days. Flag any fatigue signals and list the ads driving most of the spend.
Be patient with early data. Meta's learning phase means the first days are noisy, so avoid judging or cutting ads before the campaign has had time to stabilize.
Rules work well when they come with limits. For example:
Propose pausing any ad with CPA above $Z after at least [N] conversions or $[amount] spend. List the ads that qualify and wait for my approval before acting.
Asking Claude to propose before it executes keeps you in control while still removing the manual digging.
Close the loop on creative. When a new ad wins, make sure it lands in your creative library, then ask for the next round:
Based on the highest-scoring elements from this campaign, propose the next set of variations. Keep the winning hook and test three new visual treatments.
If you use Cometly for attribution, ask Claude to compare Meta-reported results with Cometly's conversion data before you scale or cut anything. When the two sources agree, act with confidence. When they diverge, investigate before moving budget.
Turning your first launch into a repeatable weekly workflow
Once your first campaign is live, check it against Ads Manager and confirm the structure, spend, and ads are what you approved. Then save the prompts that worked, including the analysis query, the brief template, the variation plan request, and the approval wording, as reusable templates. Each week, run the same review, feed winners back into your library, and ask for the next round of variations. The more consistent your prompts, the more comparable your results become.
