SimilarWeb MCP: AI-Powered Competitive Research (2026)

TL;DR + what you actually need

Four constants if you’re here for the snippet:

• Endpoint: https://mcp.similarweb.com/ — Streamable HTTP, hosted by SimilarWeb.
• Auth header: api-key (not Bearer / not Authorization). Pass your SimilarWeb API key as the value.
• Pricing prerequisite: a SimilarWeb subscription with API access enabled. There is no self-serve free tier for the API.
• Rate limit: 10 requests per second per account. Data credits debit per tool call.

The fast install paths, in case that’s all you came for:

• Claude Code: claude mcp add --transport http similarweb https://mcp.similarweb.com/ --header "api-key: YOUR_KEY"
• Cursor: drop the JSON block (below) into ~/.cursor/mcp.json with headers.api-key set.
• Codex CLI: a [mcp_servers.similarweb] TOML block in ~/.codex/config.toml — snippet below.
• OpenCode: add a remote MCP entry with headers in your OpenCode config — snippet below.

The rest of this guide explains why those snippets are shaped that way, walks each of the twenty-odd tools, shows six real competitive-research recipes, and covers the rate limits and credit accounting that bite first-time users on day two.

What SimilarWeb MCP gives you

Most LLM-driven competitive-research workflows fail in the same spot: the model has stale knowledge about competitor brands, can’t actually see traffic shapes, and ends up writing plausible-sounding fiction. Hooking up a panel-data provider — SimilarWeb being the most widely-licensed of them — fixes that class of bug by giving the agent a primary-source tool to call before it speaks.

The SimilarWeb MCP server is a thin MCP wrapper over the same Digital Intelligence REST API that powers the SimilarWeb web dashboard. Every visible chart in SimilarWeb — the country-rank trend on a domain overview, the traffic-source pie, the audience-overlap heatmap, the top apps in a category — comes from a REST endpoint. The MCP server exposes those endpoints as MCP tools, one per dataset, so an agent can ask the same questions a human analyst would ask the dashboard.

Practical difference: instead of you opening similarweb.com, typing in a competitor domain, clicking Traffic Sources, screenshotting the chart, and pasting it into a doc, you ask the agent “summarise competitor.com’s traffic sources for the last 90 days versus our domain.” The model fetches both, computes the delta, and writes the analysis. The data is the same; the loop is twenty times faster, and the output is grounded in real numbers instead of vibes. The wider implication is that the unit of competitive research stops being “a dashboard session” and becomes “a prompt.” A weekly competitor sweep that used to take a strategy analyst two hours turns into a five-minute prompt and a one- page summary. The first month after teams wire this up we typically see the number of competitor write-ups triple — not because anyone cares more, but because the marginal cost of one more analysis falls to roughly zero.

Worth saying plainly: the SimilarWeb MCP is a thin shim. It doesn’t synthesise insight; it gives the model access to the same numbers a human analyst would pull. The leap is that the model can chain six of those pulls together, reconcile the answers, and write the paragraph. That’s the unlock.

The official endpoint lives at mcp.similarweb.com and the canonical reference is the developer doc at developers.similarweb.com/docs/similarweb-mcp. The MCP catalog entry for the server — install buttons, transport configs, manifest — is at mcp.directory/servers/similarweb.

API setup + plan tiers

This is the section that catches everyone the first time, because SimilarWeb’s pricing model is deliberately opaque. Here’s the practical summary.

You need a SimilarWeb plan with API access. The MCP server is free in the sense that you don’t pay extra to use it — it’s a wrapper, not a separate product — but the API it wraps requires an active SimilarWeb subscription that includes API access. As of writing, that means one of three commercial paths:

• API-only standalone: SimilarWeb sells the API as its own product without a full platform seat. This is the cheapest path if you want programmatic access and don’t need the dashboard UI. Quotes come from sales, not a pricing page.
• Business package: a full SimilarWeb seat with the API added as a customisable line item. Most enterprise customers land here. API access is not on by default; you have to confirm it when you sign.
• Enterprise: custom contract, larger credit pool, the Batch API for million-row exports, SSO, dedicated support.

There is no free developer tier. Trial keys exist for evaluation but are gated through sales. If you’re doing this evaluation as a developer at a company that already pays SimilarWeb, the fastest path is to ask your account owner to enable API access on the existing seat and generate a key from the developer dashboard. Self-serve doesn’t exist.

Generating the key. Once your account has API access, log into the SimilarWeb user dashboard, open the API section (the URL is shown in the developer docs once you’re authenticated), and create a key. The key is an opaque string — treat it like any other production credential, scope it to the smallest set of services that need it, and rotate when you can.

Data credits. SimilarWeb meters API usage in data credits rather than raw request count. A simple endpoint (e.g. website rank for one domain over one month) consumes a few credits; a heavy endpoint (audience overlap across five domains over a year) consumes proportionally more. The MCP server inherits this billing model exactly — one MCP tool call = one REST call = credits debited. Composite questions that the agent answers by chaining tools obviously cost more than single-tool answers. Your credit pool refresh cadence depends on the contract; check the dashboard for the current balance and refresh date.

What that means for agent design. If you’re building a long-running agent that hits SimilarWeb hundreds of times a day, set a budget cap in the agent’s system prompt. We’ve watched a single overnight run consume a multi-thousand-credit pool because the agent looped over every competitor and every report type without a stop condition. Treat SimilarWeb tools as expensive and prompt the agent to ask for them sparingly.

The practical pattern that works: write the agent’s system prompt with two explicit rules. First, “use SimilarWeb tools only when the question requires live traffic data; for analysis of cached results, reuse the previous response in the conversation.” Second, “limit yourself to at most six tool calls per user message unless the user explicitly says ‘deep research’.” Those two rules cut credit burn by an order of magnitude without hurting answer quality. The agent learns to summarise from data it already has instead of re- fetching, and it doesn’t accidentally spin up fifty tool calls when one would do.

One more design note. The SimilarWeb panel data is generally month-level granularity for free traffic splits, with day-level only on the heavier endpoints. That’s a real-world constraint to put in your agent’s prompt: if the user asks “what did competitor.com’s traffic look like yesterday,” the honest answer is “SimilarWeb doesn’t update that fast.” Better to say that than to let the model invent a daily figure from monthly averages.

Install (every client)

The shape is the same across every client: a remote MCP server pointed at https://mcp.similarweb.com/ with a custom HTTP header. What differs is the config syntax each client expects. The install panel below pulls the exact JSON / TOML / one-line command from our catalog — tap your client’s row, copy the snippet, paste it where the panel says, drop your API key in, restart. SimilarWeb’s tools are available in that client within seconds.

One-line install · SimilarWeb

Open server page

Install

Cursor

One-click install

1-Click

VS Code

One-click install

1-Click

Claude Desktop

Copy configuration

Claude Code

Copy CLI command

CLI

Gemini

Copy CLI command

CLI

Codex

Copy CLI command

CLI

Windsurf

Copy configuration

ChatGPT

Add via Settings → Connectors

M

Manual

Copy configuration

Add SimilarWeb to Claude Code

The Claude Code one-liner is the most copy-pasted of the bunch:

claude mcp add --transport http similarweb https://mcp.similarweb.com/ --header "api-key: YOUR_KEY"

That registers SimilarWeb at user scope — meaning every Claude Code project you open can call its tools. Add --scope project if you want the entry written to .mcp.json in the repo root so it travels with the project. Verify with claude mcp list; you should see similarweb with a connected status. If you want the install to live in a repo so colleagues inherit it, commit the .mcp.json but do not commit the key — put the key in a .env.local the team shares out-of-band, or use Claude Code’s environment-variable expansion. See our Claude Code client page for the env-var pattern.

Add SimilarWeb to Cursor

Cursor reads its MCP config from ~/.cursor/mcp.json (or the project-scoped .cursor/mcp.json). Add this block:

{
  "mcpServers": {
    "similarweb": {
      "url": "https://mcp.similarweb.com/",
      "headers": {
        "api-key": "YOUR_KEY"
      }
    }
  }
}

Restart Cursor. The SimilarWeb tools should appear in the model’s tool list. The Cursor client page covers the rest of the config options — tool disabling, request timeouts, env-var expansion for the key.

Add SimilarWeb to Codex CLI

Codex CLI reads MCP servers from ~/.codex/config.toml. The TOML shape for a remote server with custom headers is:

[mcp_servers.similarweb]
type = "http"
url = "https://mcp.similarweb.com/"

[mcp_servers.similarweb.headers]
api-key = "YOUR_KEY"

Save the file, rerun codex in a fresh shell, and SimilarWeb shows up in the model’s MCP tool list. The Codex CLI install panel above emits this exact block so you don’t have to retype it.

Add SimilarWeb to OpenCode

OpenCode (the open-source TUI agent) takes a remote MCP entry in its user config — commonly ~/.config/opencode/config.json. Run opencode config print to confirm the exact path on your install. The snippet:

{
  "mcp": {
    "similarweb": {
      "type": "remote",
      "url": "https://mcp.similarweb.com/",
      "headers": {
        "api-key": "YOUR_KEY"
      }
    }
  }
}

Restart OpenCode; SimilarWeb appears in the tool catalogue.

Browse every client and its config path at mcp.directory/clients. If your client supports MCP over Streamable HTTP with custom headers, it can run SimilarWeb — the pattern is always “remote URL + header map.”

Tools walkthrough

SimilarWeb’s MCP exposes roughly twenty tools. Each one maps to a REST endpoint in the Digital Intelligence API. You don’t need to memorise them — the agent reads the tool catalogue and picks the right one for the question — but it’s worth knowing what’s in the box so you can write good prompts. Below, the tools are grouped by family.

Website Intelligence

The bulk of the API. These tools take a domain (or a list of domains) and a time window, and return shapes of traffic.

• Traffic and Engagement — visits, unique visitors, time on site, pages per visit, bounce rate.
• Traffic Sources — share of visits from organic search, paid search, direct, email, social, display, referral. The single most useful tool for competitor diagnosis.
• Website Rank — global, country, and category rank over time. Useful for “is this competitor gaining or losing share” questions.
• Subdomains — traffic share across subdomains. Reveals whether a competitor’s growth is on their core product or a side site.
• Similar Sites — sites with overlapping audience patterns to a given domain.
• Referrals — incoming and outgoing referral domains, with traffic share.
• Audience Interests — the other websites a domain’s visitors also visit, with affinity scores.
• Audience Overlap — pairwise overlap across two to five domains. The classic “how much do our visitors also see competitor.com” tool.
• Deduplicated Audience — cross-device unique visitor counts.
• Leading Folders — top URL path prefixes by traffic share. Reveals which content buckets drive a competitor’s site.
• Popular Pages — individual highest-traffic URLs.
• PPC Spend — estimated paid search spend by channel.
• Geography — visit share by country.
• Technologies — detected tech stack (analytics, payments, framework signatures).
• Demographics — age and gender breakdown of the audience.

Search Intelligence

• SERP Players — top sites ranking for a keyword across organic, paid, and zero-click placements.
• SERP Clicks — click-volume trend over time for a keyword.

Mobile App Intelligence

• App Install Penetration — the percentage of devices in a country with the app installed.
• App Store Rank — iOS and Android category rank history.
• App Details — metadata, ratings, recent version notes.

Example agent flow: you ask “summarise stripe.com’s traffic mix and top referrals for Q1.” The agent calls traffic_sources, gets the channel breakdown, calls referrals, gets the top referring domains, and writes a paragraph citing both. Two tool calls, one paragraph, real numbers — the loop the dashboard makes you do by hand.

Recipes

Six workflows where the SimilarWeb MCP pays for itself. Each assumes the server is installed at user scope and a valid key is on the connection.

Recipe 1 — The instant competitor brief

Prompt: “Give me a 200-word brief on vendor-x.com: rank, top three traffic sources, top three referral domains, country mix, audience overlap with our domain our-product.com. Cite the percentages inline.” The agent fans out across five tools, returns one paragraph. Replaces a 20-minute manual dashboard tour with a 90-second prompt. The cited percentages are the proof.

Recipe 2 — Find the traffic-leak categories

You suspect your audience also spends time on a competitor’s product but you don’t know how much. Prompt: “use SimilarWeb audience overlap to compare our-product.com against [list of three competitors] for the last six months. Rank by overlap percent and explain the direction of leakage if possible.” The agent calls audience_overlap with the five-domain limit, returns the matrix, and writes the interpretation. Surface this in your monthly board update.

Recipe 3 — Keyword gap from SERP players

Prompt: “For each of these ten product keywords, list the top three organic ranking sites via SimilarWeb SERP players, then flag which competitors appear in more than three of those lists.” Drops a de-facto keyword gap report into the chat in a minute, without you opening a single SEO tool. Pair this with Ahrefs or Semrush for backlink depth on the candidates you actually want to rank against.

Recipe 4 — Automated monthly competitive report

Bake the call pattern into a Claude Code skill or a Codex CLI prompt so it runs on a schedule. The skill takes a competitor list and a comparison domain, walks every Website Intelligence tool, and writes a Markdown report. Run it via cron or a GitHub Action on the first Monday of every month, ship it to the team Slack. We’ve seen this pattern replace a full-time competitive analyst’s monthly tour at a mid-stage SaaS.

Recipe 5 — Pricing-page traffic diagnosis

You shipped a pricing-page redesign two weeks ago. Prompt: “Pull popular pages for our-product.com over the last 30 days, focus on the /pricing path. Compare to the prior 30 days. Has its share of total site traffic moved? If yes, by how much?” The agent calls popular_pages twice (current and prior window) and computes the delta. Useful because you don’t have to log into the dashboard to spot a quiet regression.

Recipe 6 — Tech-stack scouting

Sales engineering wants to know which CRM, payments, or analytics tools a target account uses. Prompt: “Use SimilarWeb technologies on each domain in this list of 50 prospects and tag any that use Salesforce or HubSpot.” Fifty tool calls, one CSV-shaped Markdown table. Pair with Context7 if you also want the agent to suggest which integration docs each prospect’s engineers would need to read next.

A note on prompt economy across all six recipes: the cleanest pattern is to give the agent the data shape you want back, not just the data you want fetched. “Return a table with columns: domain, top source, share of organic, audience overlap with our-product.com” runs faster than “tell me about these companies” because the model knows when to stop fetching. The dashboard tour is replaced by a tightly-scoped table.

Rate limits + credit cost

Three numbers worth burning in.

• 10 requests per second hard ceiling on the REST API, account-wide. Exceeding it returns 429 Too Many Requests. The MCP server doesn’t paper over this — if your agent parallelises tool calls aggressively, you’ll hit the limit. Most clients serialise tool calls anyway, so this is rarely a problem in single-agent use; it bites when you run a fleet of agents from the same key.
• Data credits per call. Each tool invocation debits credits from your subscription pool. Simple endpoints (a single-domain rank query) cost few credits; heavy endpoints (multi-domain audience overlap over a year) cost more. The exact multiplier is in your contract.
• Composite questions multiply costs. An agent answering “summarise this competitor” might call six tools. That’s six debits. Budget accordingly — set a stop condition in the agent’s instructions if you let it loop over a list of domains.

For genuinely bulk research (millions of domains, one report type), the SimilarWeb Batch API is the right tool, not the MCP server. The MCP is a real-time conversational interface; the Batch API is an offline pipeline. Mixing them is fine: prototype with the MCP, productionise the report with Batch once you know what you want.

Troubleshooting

When to switch or pair

SimilarWeb is strongest at panel-data traffic intelligence — it knows where the world’s traffic flows, who shares audiences, what apps people install, what countries each domain serves. It is not the right tool for every research job.

• You need backlink graphs: use Ahrefs. SimilarWeb’s referral data is traffic-weighted; Ahrefs’ is link-weighted. Different question, different tool. similarweb alternatives search results lead here often — the truth is Ahrefs and SimilarWeb solve different problems and most serious competitive teams license both.
• You need keyword position depth and content gap analysis: Semrush goes deeper than SimilarWeb on per-keyword position history and SERP feature breakdowns. SimilarWeb gives you the players; Semrush gives you the long tail.
• You only need basic “is this a real company” signals: free traffic estimators (Cloudflare Radar, public domain-rank tools) give a rough indication for free. SimilarWeb is the premium upgrade once the question matters enough to bill for.
• You want to query developer documentation, not traffic: entirely different category — see our Context7 MCP complete guide for the canonical docs-RAG MCP server.

In practice we see the strongest results from teams that run SimilarWeb alongside Ahrefs (for backlinks) and Semrush (for SEO position depth), all as MCP servers in the same Claude or Cursor session. The agent picks the right tool for the question, and the analyst gets a one-prompt answer instead of three browser tabs.

The right way to think about it: SimilarWeb is your traffic layer, Ahrefs is your link layer, Semrush is your keyword layer. Each one is a partial view of the same competitor. Running them as parallel MCP servers lets a single agent prompt reconcile all three. The quality of the answer scales with the number of primary-source tools the agent can reach, not with the size of the model.

FAQ

Sources

• Official MCP doc: developers.similarweb.com/docs/similarweb-mcp
• Rate limit reference: developers.similarweb.com/docs/rate-limit (10 RPS, 429 on overflow)
• Hosted endpoint: mcp.similarweb.com
• Pricing landing: similarweb.com/corp/pricing (sales-quoted; API tier confirmed in the docs)
• Catalog entry: mcp.directory/servers/similarweb