An MCP server lets an AI assistant do real work on your mobile app instead of just talking about it. The Model Context Protocol (MCP) is a standard way to give assistants like Claude and Cursor a catalog of tools they can call. Connect once, and the chat you already use can draft an onboarding flow, publish it to your live app, answer funnel questions in plain language, start an experiment, and translate screens into new languages. Every action runs under your own account and your role's permissions, and you direct each step.
What is the Model Context Protocol?
MCP is an open standard that defines how AI assistants discover and call external tools. A tool is a typed function with a name, a description, and an input schema: think publish_flow, create_experiment, or execute_analytics_query. When you connect an MCP server, your assistant reads the tool catalog and picks the tool that fits whatever you typed. No copy-pasting API docs into the chat, no glue scripts to maintain.
Servers can run locally, but the convenient pattern for app teams is a hosted one. Setgreet's MCP server lives at mcp.setgreet.com: nothing to install, no local process, no API key to manage. Documented clients include Claude Desktop, Claude Code, Cursor, VS Code, Windsurf, and Codex CLI, plus any spec-compliant MCP client.
What does the loop look like for a mobile app?
Here is the loop a connected assistant can run end to end, using Setgreet's tool catalog as the worked example. The subject is the thing your growth actually depends on: the onboarding flow, the paywall, the surveys. Each step is a single chat message.
- Compose. Describe the flow you want, in your words. The compose_flow tool asks a short set of intake questions, proposes a plan of screens, and generates a draft made of real, editable native screens. Generating a flow spends an AI credit. Nothing is live yet.
- Publish. When the draft looks right, publish_flow pushes it to your live app without an App Store release. Even a fast review is still a cycle: Apple reports 90% of submissions are reviewed in under 24 hours, which still means waiting a day to fix a line of onboarding copy.
- Measure. get_flow_dashboard and get_flow_paths return completion rates and drop-off by screen. For anything off the beaten path, ask in plain language: execute_analytics_query turns the question into a typed, org-scoped query, not raw SQL.
- Experiment. create_experiment sets up variants, and get_experiment_results reports per-variant numbers with statistical significance.
- Translate. bulk_translate_flow localizes every screen while preserving {{personalization}} variables, so "Welcome back, {{user.firstName}}" survives in every language.
The volume of iteration this unlocks is the point. An analysis published by Adapty in 2026 found that apps running 50 or more experiments had 18.7x the median revenue of apps running one. That is correlation, not causation, but the direction is hard to ignore: teams that test more learn faster, and lowering the cost of each test is exactly what a tool loop does. The compose and edit tools sit on the same engine as Build with AI in the dashboard, and experiments work the same way in both places.
How do you connect?
Add https://mcp.setgreet.com/mcp to your client's MCP settings. The first tool call opens a browser window where you sign in with your existing Setgreet account: no API key is generated, copied, or pasted anywhere. Under the hood this is OAuth 2.1 with PKCE. Access tokens last five minutes, refresh tokens rotate and expire after 30 days, and every session is listed in Settings under MCP Connections, where you can revoke any client at any time. Per-client setup guides live in the MCP docs.
What stays in your control?
Permissions come from your dashboard role, not from whatever the AI client requests. A viewer gets read access; editing and publishing require member access or above. A client cannot talk its way into scopes your role does not grant.
Destructive operations, such as deleting a flow, unpublishing, or completing an experiment, require an explicit confirmation before the server will execute them. Chat edits target the draft, never the published version, and the flow editor's canvas hot-reloads so you can watch each change land. If two edits collide, the second one returns a conflict instead of silently overwriting the first.
Costs are visible too. Generating a flow spends AI credits, editing by chat is free, and your balance shows in the flow editor and in Settings under Billing.
When does MCP beat the dashboard?
MCP wins wherever the work is repetitive or question-shaped. Renaming a CTA across every screen, translating a flow into five languages, or rerunning last week's funnel check are one message each instead of a run of clicks. It also wins for analytics questions you would otherwise build a chart to answer: "where do users drop off?" is faster to type than to configure. And it wins for chained work: "duplicate the flow, swap the paywall screen, start a 50/50 experiment" is a single request.
The visual editor wins at visual work. Spacing, imagery, color, and the feel of a screen on an actual device are things you judge with your eyes, and a chat transcript cannot show you that. Fine-tuning a layout by describing it is slower than dragging it.
The two are not in competition. Chat edits and editor edits land in the same draft, so a common afternoon looks like: compose in chat, polish in the editor, publish, then ask the assistant how the funnel moved.
Frequently asked questions
Do I need to install anything to use Setgreet's MCP server?
No. It is hosted at https://mcp.setgreet.com/mcp over Streamable HTTP, so you add the URL to your client's MCP settings and sign in through the browser when prompted. Claude Desktop, Claude Code, Cursor, VS Code, Windsurf, and Codex CLI are all documented clients. ChatGPT support exists but is in beta with plan restrictions, so do not expect full parity there yet.
Can the AI publish or delete things without me?
The assistant only acts when you send it a request, and always within the scopes your dashboard role grants. Destructive operations like deleting a flow, unpublishing, or completing an experiment require explicit confirmation before they run. Publishing executes when you ask for it, so treat a publish request as the deliberate step it is, and you can revoke any connected client from Settings under MCP Connections at any time.
Does using the MCP server cost extra?
Generating a flow with compose_flow spends AI credits, editing by chat is free, and translation spends credits proportional to the work. Your current balance is always visible in Settings under Billing, and top-up packs are available if you run out mid-sprint. For plan details, see pricing.