Let your AI agent read your app reviews: an MCP quickstart

In short: Tritonium runs a remote MCP server for your App Store and Google Play reviews. Add one URL to your AI agent, sign in through your browser, and the agent can search reviews, read insights and forecasts, queue an analysis, and draft replies that wait for your approval. This quickstart uses the same steps we test against production with Claude Code, Codex, Cursor, GitHub Copilot CLI, Grok CLI, Antigravity, and OpenClaw.

What MCP is

The Model Context Protocol (MCP) is an open way for AI agents to call tools on external services. Instead of exporting CSVs and pasting them into a chat, your agent asks the service directly: which reviews came in, what themes stand out, what the rating forecast looks like. If you already use an MCP-capable agent for code or docs, app reviews become one more tool it can reach.

Tritonium's server speaks Streamable HTTP with JSON responses at one endpoint: https://api.tritonium.com/mcp. It implements the MCP authorization flow (OAuth 2.1 with PKCE), so the recommended setup stores no API key anywhere.

Connect your agent

Every supported client needs only the server URL to start. The first connection opens Tritonium in your browser: sign in, then approve exactly the permissions the agent gets. Read scopes (apps, reviews, analytics, alerts) are available to every workspace member; write scopes (reply drafts, analysis runs, webhooks) need an admin or owner.

claude mcp add --transport http --scope user tritonium https://api.tritonium.com/mcp

{"mcpServers": {"tritonium": {"url": "https://api.tritonium.com/mcp"}}}

codex mcp add tritonium --url https://api.tritonium.com/mcp

What the agent can and cannot do

The agent acts as a workspace member limited to the scopes you approved. It can:

• Read apps, reviews, AI insights, rating history and forecasts, alerts, webhooks, and workspace activity.
• Queue an AI analysis for a connected app and follow or cancel the run.
• Write reply drafts. Drafts are never published automatically, by anyone, agent or otherwise.

It cannot approve or publish replies, connect or remove apps, change alert rules, manage billing or team settings, or create webhooks. Those stay with signed-in admins and owners in the dashboard, so an agent session can never become a publishing channel on its own.

Three prompts to try

• "What are the top complaints about my app this week?" The agent searches recent reviews and reads the current insights, then points at the source reviews behind each theme. Check a few before acting: a summary is a starting point, not proof.
• "Queue an analysis of my latest reviews and tell me what changed." The agent starts an analysis run, follows it, and summarizes the new themes against what you already had.
• "Draft a reply to this 1-star review." The agent writes a draft that lands in your approval queue. Review the tone and the facts, edit if needed, and approve it yourself in the dashboard.

Safety and limits

Sign-in tokens last one hour and renew automatically while the agent is in use; after 30 days idle you sign in again. Every connected agent is listed under Settings, and disconnecting one revokes its access without touching your password. If you run agents in headless scripts, a scoped API key works instead of sign-in; keep it in an environment variable, not in a config file.

Two honest limits: analysis starts with your newest reviews and adds older ones over time within each app's monthly AI allowance, so answers about long history can be partial (the coverage figure is shown with each insight). And rating forecasts are model estimates, so treat "what will my rating be" answers as directional, not certain.