An SEO I follow on LinkedIn, Anna York, recently posted that Claude had just replaced 30 minutes of her Semrush dashboard work with one prompt. In plain English. She asked Claude to list the top domains in the German dog food market for January 2026, rank them by organic traffic, then show who was growing fastest year over year. The answer came back in seconds. Her post reached 54,000 people in two days.
That prompt was only possible because of Semrush MCP, the official connector that plugs live Semrush data directly into AI tools like Claude, ChatGPT, Cursor, and VS Code. No CSV exports. No copy-paste. Semrush product marketer Paloma Gracia calls it a bridge between an AI tool and a live data source, and that framing is accurate. The underlying standard, Model Context Protocol, comes from Anthropic.
What this guide covers:
- What data Semrush MCP actually exposes
- The two-minute Claude Desktop setup
- Which plans include it
- What it costs in API units
- The prompts I keep reusing
- Where the Semrush web UI is still faster
What MCP Actually Is (and Why Semrush Built One)
Model Context Protocol is an open standard Anthropic released in late 2024, and the easiest analogy is USB-C for AI tools. One standard port, any compatible data source plugs in.
Before vs. after MCP
| Before MCP | With MCP | |
|---|---|---|
| Getting data into AI | Export CSV, paste into prompt, hope the AI understands the columns | AI queries Semrush directly |
| Alternatives | Build a custom plugin, or point ChatGPT’s browser tool at a Semrush page and pray | Direct, real-time connection |
| Workflow | Fragmented, manual | AI queries, Semrush responds, AI turns it into analysis |
Launch and status
- Announced: September 10, 2025
- Consumer-facing explainer: February 10, 2026
- Endpoint:
https://mcp.semrush.com/v1/mcp - Compliance: SOC 2 Type II
This is not a beta, not a community side project, not vaporware. It’s a shipping, officially supported product, which I mention only because one competing article I read still hedged with a “cannot be confirmed” disclaimer. It can be confirmed. I’ve connected it.
Why it matters for SEOs: compression
Why it matters for SEOs specifically is compression. I used to export a keyword gap CSV, paste it into Claude, then ask a follow-up and realise I needed a fresh export. Now Claude pulls the numbers itself. Zaki Akhtar put it well in the comments on Anna York’s post: “The biggest change isn’t automation. It’s compression of research time.”
So what can Claude actually see when it’s connected? That’s the part the marketing page skims over.
What Semrush MCP Actually Exposes: Data, Tools, and Limits
The underlying data pool
The scale Claude is reaching into every time you ask it a question:
- 27.3 billion keywords
- 808 million domains
- 43 trillion backlinks
- 142 geographical databases
The three API surfaces
Semrush splits MCP access into three API surfaces, and the distinction matters because they’re billed differently:
| API surface | What it covers | Plan requirement |
|---|---|---|
| Standard API | Keyword research, domain overview, organic and paid competitors, backlinks, phrase questions | Included with Pro, Guru, and Business plans |
| Trends API | Traffic analytics, market share, geographic traffic breakdowns, competitor benchmarking | Requires a separate Trends subscription |
| Projects API | Read-only access to your own Semrush projects, site audit results, and position tracking data | You can query your setup; you cannot create or modify it |
Tool names, translated
The tool names Semrush uses are descriptive but technical. Here’s how I translate them when writing prompts:
| Tool name | What it means in plain English |
|---|---|
semrush_domain_overview | “give me the snapshot for competitor.com” |
semrush_competitors | “who competes with this domain organically?” |
semrush_related_keywords | “expand this seed keyword” |
semrush_keyword_difficulty | “how hard is it to rank for X?” |
semrush_keyword_overview | “volume, CPC, and intent for this term” |
semrush_phrase_questions | “what questions are people asking about this?” |
semrush_backlinks | “top backlinks pointing to this URL” |
semrush_traffic_summary | “how much traffic does this site get?” (Trends required) |
semrush_domain_organic_keywords | “what terms does this domain already rank for?” |
semrush_api_units_balance | “how many API units do I have left?” |
You don’t need to memorise any of this. Claude picks the right tool when you ask the question in plain English. The names only matter when you want to predict what Claude is about to call.
One honest boundary: read-only
The MCP is read-only across all three surfaces. It cannot:
- Create a project
- Trigger a site audit
- Publish content
- Change a position tracking campaign
If you need the write side of Semrush, you still have to click into the dashboard.
Setting Up Semrush MCP: Claude, ChatGPT, and Developer Tools
If you use Claude Desktop, you can have Semrush MCP connected before your coffee finishes brewing. Three clicks and an OAuth approval. Here’s the path for each client I’ve tested.
Claude Desktop (no-code, most common)
- Open Claude, go to Settings
- Click Connectors in the sidebar
- Click Add custom connector
- Name it Semrush
- Paste the URL
https://mcp.semrush.com/v1/mcp - Save
Claude redirects you to a Semrush OAuth screen. Approve. Done. No JSON, no terminal, no config file.
ChatGPT (also no-code, different pattern)
ChatGPT doesn’t expose raw MCP endpoints to users. It uses a dedicated Semrush app instead.
- Go to Settings → Account → Apps
- Find Semrush
- Click Connect → Continue → Approve
- Tag
@Semrushat the start of any prompt where you want it to pull data
Same end result, different shape under the hood.
Claude Code (one command)
If you use the Claude Code CLI, the entire setup is:
claude mcp add semrush https://mcp.semrush.com/v1/mcp -t http
Then run /mcp inside Claude Code, pick Semrush, and approve the OAuth flow.
Cursor and VS Code (small JSON snippet)
Both tools use a config file. The Cursor version goes under Settings → Tools and MCP:
{"mcpServers": {"semrush": {"url": "https://mcp.semrush.com/v1/mcp"}}}
VS Code uses a workspace-level .vscode/mcp.json file with a near-identical snippet:
{"servers": {"semrush": {"url": "https://mcp.semrush.com/v1/mcp", "type": "http"}}}
Gemini CLI is the odd one out. It needs the API key header method, not OAuth.
Hosted vs self-hosted
Most readers want the hosted version I just walked through. Two community GitHub alternatives exist:
mrkooblu/semrush-mcp— exposes 77 tools across 8 categories (wider coverage than the official server lists by fixed name)metehan777/semrush-mcp— a TypeScript implementation for developers building custom integrations
Both self-host via npm install and a local .env file. Skip this path unless you need:
- Broader tool coverage
- Local caching
- Custom rate limiting
Do you need a developer? No for the official hosted version. A little bit yes if you go the self-hosted route.
Plans, API Units, and What Semrush MCP Really Costs
The sentence nobody at Semrush wants to lead with: MCP runs on the same API units as the regular Semrush API, and it’s entirely possible to burn through a month’s allocation in an afternoon if you let Claude loose on historical data.
Which plans include it
MCP access ships with the following plans, each coming with 50,000 Standard API units per month:
- Semrush One (any tier)
- SEO Classic Pro
- SEO Classic Guru
- Business
For traffic data you also need Trends API Basic or Premium. Free plans and entry-level tiers are out.
The real combined cost
Starting from scratch, a useful setup is:
| Tool | Monthly cost |
|---|---|
| Semrush Pro | $139.95 |
| Claude Pro | $20 |
| Total | ~$160/month |
If you already pay for Semrush, the AI side is marginal. If you don’t, the AI tool sits on top of the subscription, it doesn’t replace it.
What a query actually costs
Semrush doesn’t publish a definitive per-tool unit table. Based on Gumloop’s credit mapping and Semrush’s own API pricing docs, here’s what I work from:
| Query | Approximate unit cost |
|---|---|
| Domain rank report | ~4 units |
| Domain overview | ~5 units |
| Keyword overview | ~5 units |
| Keyword difficulty | ~8 units per call |
| Backlinks historical | ~10 units per call |
| Pulling 1,000 organic keywords live | ~10,000 units |
| Pulling those same 1,000 keywords historically | ~50,000 units |
That last line is the trap. Historical data costs roughly five times what live data costs. One enthusiastic historical backlinks lookup on a big domain can erase your whole monthly budget. Agentic loops that pull historical endpoints on a schedule will quietly empty your account while you sleep.
Three habits that conserve units
- Ask Claude to run
semrush_api_units_balancebefore any big job - Prefer regional queries over global (global costs more)
- Never let a scheduled workflow touch historical endpoints without a hard unit cap
Real Workflows: Prompts Marketers Are Actually Using
The hardest part isn’t setup, it’s knowing what to ask first. Here are five prompts I keep coming back to, each mapped to the tool Claude ends up calling.
1. Keyword gap
“Identify high-intent keywords that competitor.com ranks for but my-site.com does not. Filter to difficulty under 40 and search volume over 500.”
Hits semrush_domain_organic_keywords on both domains and diffs them.
2. Market mapping
“List the top 10 domains in the [industry] space in [country] for [month]. Include estimated organic traffic and rank them. Then tell me which is growing fastest year over year.”
Anna York’s prompt, almost verbatim.
3. Content brief enrichment
“Use
semrush_phrase_questionsto pull the top questions people ask about [seed keyword], then give me the top 20 related keywords grouped by intent.”
4. Backlink scan
“Show me the top 50 referring domains to competitor.com from the last 90 days, flagged by authority score, and tell me which ones also link to my-site.com.”
5. Ranking risk check
“Which tracked keywords in the [project name] project dropped more than three positions this week, and who overtook me?”
Runs through the Projects API, read-only.
The follow-up is the real unlock
The bigger unlock isn’t the first prompt. It’s the follow-up. Anna York’s breakthrough wasn’t the list of dog food domains. It was typing “Who’s growing fastest YoY?” right after. Claude already had the context, so the second query was effectively free. That’s the compression Zaki Akhtar meant.
Stacking MCPs
Claude supports multiple MCP connections at once, and this is where things get genuinely interesting. Connect Semrush, Google Search Console, and GA4 MCPs in the same session and you can run prompts no single dashboard allows. The one I steal constantly:
“Which pages rank positions 1 to 5 in Search Console, have a bounce rate over 60% in GA4, and are losing organic visibility to competitors per Semrush?”
Victoria Kurichenko used a similar stack to build a 13-page SEO content strategy aimed at doubling revenue from $40K to $100K. Cross-tool reasoning like this is the main reason I bothered connecting MCPs at all.
Where Semrush MCP Falls Short: Limitations and Gotchas
Here are the things I wish someone had told me before I connected Claude to my Semrush account.
Agentic workflows drain units fast
Let an agent run on a schedule, especially one that touches historical endpoints, and you’ll wake up to a zeroed-out balance. Always wrap scheduled jobs with a pre-flight semrush_api_units_balance check and a spending cap you set yourself.
AI responses can be wrong
Semrush’s own developer docs carry the disclaimer that AI responses may be incomplete or inaccurate. Victoria Kurichenko put it plainly: hallucinations can still happen even when the AI is wired to real data. Verify any number before you put it in a client deck. The MCP reduces hallucination risk, it doesn’t eliminate it.
The read-only ceiling is real
Claude cannot create a Semrush project, trigger a site audit, update a position tracking campaign, or change anything else in your account. Write operations still require the Semrush UI.
Trends API is a separate bill
If you ask Claude for traffic analytics and it comes back empty, the most likely reason isn’t that the query failed. It’s that your plan doesn’t include a Trends subscription. Traffic summary, audience insights, and geo distribution all require the Trends tier.
The web UI is still faster for dense audits
For multi-competitor comparisons with lots of filters toggling at once, the dashboard beats narrating the same thing to an AI.
- Use the UI when you’re exploring
- Use MCP when the question is clear and you want structured output
Semrush MCP FAQ
When did Semrush MCP actually launch?
Announced September 10, 2025, and listed on PulseMCP as officially released on October 7, 2025. Paloma Gracia’s consumer-facing blog explainer went up February 10, 2026, which is why some older roundups still claim it doesn’t exist.
Do I need a developer to connect it?
- No on Claude Desktop, ChatGPT, or Claude Code. The official hosted MCP uses OAuth and a URL field.
- A little yes if you self-host the GitHub version (Node.js 18+,
npm install,.envfile).
Which Semrush plans include MCP?
Semrush One (any tier), SEO Classic Pro, SEO Classic Guru, Business, and Trends API Basic or Premium for traffic data. Free and entry-level plans are excluded. Every qualifying plan includes 50,000 Standard API units per month.
Can I use it with ChatGPT, or only Claude?
Yes, ChatGPT works. It uses a dedicated Semrush app you tag as @Semrush in prompts. The MCP is also supported in:
- Cursor
- VS Code
- Gemini CLI
- Perplexity
- Windsurf
- Codex CLI
What happens if my API units run out mid-session?
Semrush returns an error response and Claude surfaces it in plain text. It doesn’t silently stop or hallucinate data to fill the gap. Buy a top-up pack, or wait for the next monthly reset.
Is the Trends API included?
No. Trends is a separate subscription. Traffic summary, audience data, geo distribution, and competitor traffic benchmarking all require Trends Basic or Premium on top of your SEO plan.
Comments 0 Responses