screenroveMCP integration examples: how SaaS apps connect to coding agents
MCP integration examples from real SaaS screens: Eraser's claude mcp add command, Amplitude's Cursor prompt, Tally's AI-first grid and Linktree's locked MCP.
In 2026 the integration a SaaS product asks you to set up first is less often Slack or Google Sheets and more often your AI coding agent. On our integration pages wall, 8 of the 48 products (9 of 83 screens) offer to connect the product to an outside AI client such as Claude, Cursor or ChatGPT. At Amplitude, Tally, Linktree and Lucid, that option is listed first. These are the MCP integration examples we captured, and they share a small set of mechanics worth copying.
The most direct is Eraser. Its onboarding has a step called "Connect your MCP client" that shows no OAuth button, only ten client tiles and a command to paste.
MCP integration examples, by mechanic
Across the screens, the connect step takes one of three forms: a command, a prompt or a client picker. Most products combine two.
1. The command: one line to paste in a terminal
Eraser's ten tiles are Claude Code, Claude.ai, ChatGPT, Codex, Gemini, OpenCode, VS Code, Cursor, Kiro and GitHub Copilot. Choosing Claude Code fills a card with claude mcp add --transport http eraser https://app.eraser.io/api/mcp and a Copy chip. There is no API key to generate and nothing to fill in. Eraser reuses the same picker in MCP settings, and "Connect Eraser MCP" is one of the five quick-start tiles on the empty dashboard, so a user who skips it in onboarding sees it twice more.
Hotjar does the same for its tracking tag. The install modal (headed "Install Contentsquare on your site") opens on an "AI-assisted" tab ahead of Manual, Tag Managers and Platforms, and its default Wizard option is a single npx @contentsquare/wizard install command that "configures your AI agent, installs the tag for your framework, and verifies it's working." It ships with a disclaimer: Contentsquare "is not responsible for how your tool interprets or applies the wizard instructions." PostHog puts a similar npx -y @posthog/wizard@latest warehouse box above its 912 data sources.
2. The prompt: setup written for an agent to read
The second form skips the protocol and hands the user text for their agent. Amplitude's "Let's get data in" page leads with an AI prompt card tagged Recommended: "Run this custom prompt in an AI coding tool like Cursor or Claude to set up Amplitude." The prompt itself is one goal, not a tutorial: "one real, app-produced first event, ingested into the selected project." Two buttons follow, "Open in Cursor" and "Copy Prompt".
The interesting part is what moved down. Amplitude's integrations card ("Segment, Snowflake, and 100+ sources") now sits beneath an "or set up another way" divider, next to "Invite Team: Hand setup to a developer on your team". The agent prompt has replaced the developer as the default way in.
GitBook ends its onboarding the same way. Step 4/4, "Set up your local workflow", offers Claude, Cursor and ChatGPT as buttons, then an "or" and a Markdown prompt that opens "# GitBook Setup Agent / ## Goal / Get my docs live on a published GitBook site."
Note that the prompt has a definition of done ("Done = you…" is visible before the fade). Both Amplitude and GitBook write their prompts as an outcome the agent can check, not a list of steps.
3. The client picker: logos as the interface
The third form is a grid of AI clients where you would expect a grid of apps. Zapier has a dedicated "Create MCP server" page ("Connect your AI agent to Gmail, Slack, Salesforce, and 9,000+ more apps") with Popular tiles for Claude, Claude Code, ChatGPT, Cursor and OpenClaw, and MCP is one of five "Start from scratch" cards on its home, alongside Zap, Agent, Table and Form.
Tally folds AI clients into its normal directory, and puts them first.
Lucid goes further and gives MCP its own taxonomy. The integrations directory lists "AI" and "MCP Connectors" as the first two categories, and ships a separate listing per client: "Lucid MCP for Cursor", "Lucid MCP for Github", "Lucid MCP for Glean", each with a New badge. The app header carries a "Connect AI tools" button, and the New menu lists "Connect your AI tools / Lucid MCP" next to Lucidchart and Lucidspark. Linear's public integrations page has "Agents" and "AI clients" directly under Essentials in its category rail.
Where the agent entry point sits
The placement says how much a product is betting on it. We saw four positions:
- An onboarding step. Eraser (step 4 of 5), Amplitude (the data step) and GitBook (the last step). Each has a skip or an alternative, but the agent path is the default.
- The dashboard. 8 of the 93 products on our dashboards wall put an outside-agent entry on the home screen: Eraser, ToolJet, Vercel, Zapier, Lucid, Make ("Build and run scenarios from Claude or ChatGPT"), Softr ("Use Softr from" with three AI logos in the workspace menu) and Krea ("Get Krea into ChatGPT").
- The top of settings. Linktree.
- The marketing hero. Clerk, Rive and Netlify, before sign-up.
ToolJet's home is the clearest dashboard example: its own AI builder sits on top, and directly below it the product offers to step aside.
ToolJet's sign-up page makes the same argument as a diagram, wiring Cursor, Claude Code and Codex into "ToolJet MCP". Vercel's empty projects overview puts an "Onboard your coding agent" pill, with three agent logos and a copy icon, between "Deploy your first project" and the Import Project row. On Netlify's new project screen, "Cursor Origin" appears as a Git provider next to GitHub and GitLab.
On the marketing hero, before an account exists
Three landing pages we captured put an agent instruction in the hero with a copy button. Clerk has a "Build with agents" box under its headline reading Add Clerk auth to my app: clerk.com/SKILL.md. Netlify puts "Add Netlify to your coding agent" under its two buttons, with fetch https://netlify.ai. Rive goes furthest: "Start with your coding agent. No account needed to try it." above a prompt box, "Install the Rive CLI from rive.app/cli. Let's build a 3D city.", with a COPY button as the main action.
MCP as a paid feature
Linktree shows the other direction this is going. In settings, "Agent access (MCP)" is the first item under Integrations, ahead of social media, and the "Jump to" select reads "Agent integrations". It promises "Let Claude, ChatGPT, or another AI tool update your links, bio, and appearance and pull your analytics", and then it is locked.
The same screen is on our paywalls wall: a feature that shipped with a New badge and a padlock together. Eraser, by contrast, offered its MCP connection during a free trial with no upgrade marker on the step. Which way to go depends on whether agent access is how users get value (give it away) or a power feature heavy users will pay for (gate it, as Linktree does).
What to steal
- Ship the exact command, not a docs link. Eraser's command includes the transport, the server name and the URL. A user can paste it without reading anything.
- Let people pick their client first. Ten tiles (Eraser) or three buttons (GitBook) both work; the command or prompt then changes to match.
- Write setup prompts as goals with a finish line. Amplitude asks for "one real, app-produced first event"; GitBook defines "Done". The agent can check its own work.
- Keep the old path one divider away. Amplitude's "or set up another way" and Hotjar's Manual and Tag Managers tabs keep non-agent users moving.
- Show it again after onboarding. Eraser repeats the picker in settings and as a dashboard tile; Lucid adds a header button and a New menu item.
- Say what the agent can do. Linktree's one line ("update your links, bio, and appearance and pull your analytics") is more useful than the word MCP.
How we counted
We captured these screens in desktop browsers with research accounts in September and October 2026. A vision model recorded each screen's attributes, which you can check on each screen's page; we counted a product when a captured screen offered to connect it to an outside AI client, so products that have MCP servers we did not see are not included. Browse the MCP screens or every connect step on the wall. We record what we saw, not A/B results, so nothing here says which approach converts better.
Screenrove is itself an MCP server: connect it and ask your agent for MCP integration examples to build your own connect step from.