Why I built an MCP server into my feedback tool

Hauke Jung

Hauke Jung

— 4 min read

Most SaaS tools have an API. Some have webhooks. Very few have a Model Context Protocol server that lets AI assistants query, filter, and act on your data conversationally.

I built one into SeggWat, a feedback widget for websites. Here's what it does, how it works, and why I think MCP is becoming table stakes for dev tools.

The real problem with feedback tools

Collection is solved. Nobody actually consumes the result.

The pattern I see all the time:

  1. A team installs a feedback widget.
  2. Users submit bug reports, feature ideas, and ratings.
  3. Feedback piles up in a dashboard nobody opens.
  4. Product decisions get made on gut feel anyway.

Dashboards make you go and look. Email digests get filtered. Slack notifications turn into noise. The signal is there, no one is in the chair.

So I started asking a different question: what if the feedback data was available wherever I'm already working, including inside my AI coding assistant?

What MCP actually is

The Model Context Protocol is an open protocol from Anthropic. A standard way for any LLM to ask structured questions to your application. Think of it as the REST API for AI workflows: instead of your own code asking questions you anticipated, any AI can ask anything and the protocol negotiates the right tool calls.

Before:

Me: "What are users complaining about?"
AI: "I don't have access to your feedback data."

With MCP:

Me: "What are users complaining about?"
AI: calls list_feedback → "47 submissions this week. Top issues: mobile login timeout (12), dark mode contrast in settings (8), broken Stripe redirect (3)..."

Now the AI is actually useful for the question you wanted to ask.

How SeggWat's MCP server works

The server lives at https://seggwat.com/mcp. It uses streamable HTTP transport: clients POST JSON-RPC, the server replies (and streams) via Server-Sent Events. Auth is your project's API key, passed either as Authorization: Bearer <key> or X-API-Key: <key>. Every session is scoped to a single project. No cross-tenant data leakage.

The flow:

Client (Claude Desktop, Cursor, etc.)
  → POST /mcp
  → JSON-RPC tool call
  → API-key auth + project scoping
  → MongoDB query
  → Structured JSON back
  → AI interprets and presents

The tools

The current tool surface:

  • list_projects: projects under your API key.
  • list_feedback, get_feedback: query and read feedback with filters for status, type, search, and page.
  • create_feedback, update_feedback, delete_feedback: full lifecycle. The AI can triage, change status, write resolution notes, soft-delete.
  • list_ratings, get_rating: query helpful, star, and NPS ratings.
  • create_rating, create_typed_rating, delete_rating: write ratings of all three types.
  • get_rating_stats, get_typed_rating_stats: aggregates including helpful percentages, star distributions, and NPS with promoter / passive / detractor breakdown.

Every tool returns structured JSON the AI can compare and reason about.

Use cases I actually run

Morning standup context. I ask "summarize feedback from the last 24 hours, flag anything critical" and get a prioritized briefing in chat instead of opening a dashboard.

Sprint planning. "What are the top requested feature ideas this month? Group by theme." The AI clusters by topic and gives me a ranked list, no manual sorting in a UI.

Release validation. The widget tags submissions with your app version via the data-version attribute, so I can ask "compare feedback between v2.3 and v2.4, did the new onboarding reduce complaints?" and get a real diff, not a feeling.

Automated triage. Pair the MCP server with a coding agent. After deploy: "any new feedback mentioning errors since 14:30 UTC? If critical, open a GitHub issue." That closes the loop between deploy, feedback, and action without anyone watching a dashboard.

Customer success. "Show me all 1-star and 2-star ratings this week, draft personalized follow-ups." The AI already has the page, version, and message, so the replies reference the actual issue.

Why not just use the REST API?

You can. SeggWat has one. The difference is intent.

A REST API answers questions your code anticipated. An MCP server answers questions someone thinks up on the spot. With REST you build integrations: GET /feedback?status=new&limit=10. With MCP you say "what's going wrong?" and the model picks the right tools and combines the results.

That gap gets bigger every month.

The Rust side

The dashboard is Rust on Axum. The MCP server runs in the same process, using the rmcp crate for transport. Tools are async methods on a SeggwatTools struct, annotated with #[tool(description = "...")], and the macro generates the JSON schema, so request validation happens at compile time. Serialization is serde. No runtime surprises.

Lessons from building it:

  • Tool names and descriptions are not decoration. list_feedback works, query confuses the model into pseudo-SQL. Be specific.
  • Keep responses structured but readable. Add computed fields like days_since_submission so the AI can reason temporally without a date library.
  • Per-project auth, not god-mode keys. MCP sessions are long-lived, and one over-permissioned key is a long-running blast radius.
  • Rate-limit like the REST API. A single prompt can fire three to five tool calls, and a curious AI can hammer your database without one.

Where this is going

The SaaS tools that survive the next few years will be the ones natively accessible to AI workflows. "Add a chatbot" is the bare minimum. Actually being AI-native looks like:

  • Exposing your data via MCP so any assistant can query it.
  • Accepting structured input so agents can take real actions.
  • Emitting events so workflows can react in real time.

A feedback tool that only works through a web dashboard loses to one that works inside the IDE, the CI pipeline, the Slack channel, and the AI assistant you already have open.

SeggWat's MCP server is a bet on that. It's already how a lot of our more technical users interact with their feedback.

Try it

  1. Sign up at seggwat.com.
  2. Drop one script tag on your site.
  3. Generate an API key in Settings → API.
  4. Point your MCP client at https://seggwat.com/mcp with that key.
  5. Ask questions.

EU-hosted, GDPR-friendly, screenshot annotation, ideas portal, and an MCP server. $6/month, 14-day free trial.

Originally published on the SeggWat blog.

Read more