mcp-server-product-studio

A full-feature MCP server for the Connext platform — a small product-development copilot. It shows how to combine, in one server:

• 🔐 Its own login — the server is its own OAuth 2.1 provider with a simple username/password page (same pattern as mcp-server-example).

• 🔎 A RAG tool — semantic-ish search over product documents (customer interviews, feature requests, PRDs, competitor notes) with a dependency-free BM25 retriever.

• 🗄️ A SQL tool — a read-only SQL query over a seeded product database (features / experiments / metrics).

• 📊 A chart MCP App — tools that render an SVG chart (in the chat), derived from the product data.

• 📝 A writable form MCP App — a feedback form the user fills in on the Connext side; on Save it calls a tool back over the app bridge and the server persists the note, then re-renders the saved list.

• 👤 Per-user identity — tools run as the signed-in user (e.g. "my features").

It's fully self-contained: a seeded in-memory SQLite database + an in-memory doc corpus, so it clones and runs with zero external services. Built on FastMCP.

What it demonstrates

The demo interactions that show them working together:

• "What are customers saying about onboarding?" → RAG

• "Show build-stage features ranked by RICE" → SQL → chart

• "How is the Onboarding checklist doing?" → SQL + adoption chart

• "Log some feedback about SSO" → form → the note is saved and shown

• "Draft next sprint's priorities" → RAG (pain points) + SQL (backlog)

Related MCP server: SAP Datasphere MCP Server

Quick start

Requires Python 3.11+.

# 1. install
python -m venv .venv && source .venv/bin/activate
pip install -e .

# 2. run
python server.py
# -> serving on http://localhost:8000  (MCP endpoint: http://localhost:8000/mcp/)

# 3. in another terminal, connect the way a client does (opens a browser login)
python examples/connect_with_client.py
# sign in as  alice / password123   (or  priya / hunter2)

Demo users are product managers whose usernames own features in the data, so "my features" works:

The tools

search_product_docs(query, limit=3) — BM25 retrieval over the doc corpus in rag.py (customer interviews, feature requests, support themes, competitor notes, PRDs). Returns the most relevant passages with scores.

describe_data() → run_sql(sql) — the model calls describe_data for the schema, then run_sql with a SELECT. The query runs against a SQLite file opened read-only (mode=ro) and is additionally checked to be a single read-only statement — a tool can never mutate the data. The signed-in username is available for WHERE owner = '<username>'.

chart_roadmap(owner=None) and chart_feature_adoption(feature) — MCP Apps. Each returns a text summary (what the model reads) and structuredContent (the chart data). Connext reads the shared ui://product-studio/chart template and its JavaScript draws the SVG from that data over the app bridge (see MCP Apps).

feedback_form() → log_feedback(feature, sentiment, note) — a writable MCP App. feedback_form opens the form with the feature list + recent notes; on Save, the form's JS calls log_feedback back over the tools/call bridge, which does a single parameterized INSERT (a read-write connection — run_sql stays read-only) and returns the updated list. Mark log_feedback app-callable in Connext so the form is allowed to call it.

The data (all seeded, self-contained)

db.py — three tables modelling a team building a SaaS product:

features(id, title, stage, priority, rice_score, effort_weeks, owner, target_release)
  stage ∈ idea → discovery → design → build → beta → ga   (NPD stage-gate)
experiments(id, feature_id, hypothesis, metric, control, variant, lift_pct, status)
metrics(feature_id, week, adoption, retention)            (weekly time series)

rag.py — ~10 short product documents (interviews, requests, PRDs, competitor notes) that reference the same features, so RAG and SQL tell a joined story.

MCP Apps: charts (read) + a form (write)

An MCP App is a ui:// HTML resource the host renders in a sandboxed iframe. Connext reads the resource (resources/read) and talks to it over the SEP-1865 JSON-RPC bridge (postMessage), so the drawing/logic lives in the template's inline JS — no external scripts or assets (strict CSP) — and it reads the host's var(--mcp-color-*) theme tokens. Each template also reports its height via ui/notifications/size-changed so the host fits the iframe to the content.

Charts (read-only) — charts.py. The tool sends the chart data as structuredContent; the host delivers it to the template (ui/notifications/tool-result) and the JS renders the SVG. The two-series colours are a validated categorical palette (blue/aqua, checked for colour-blind separation and contrast).

Form (writable) — feedback.py. Same bridge, plus the write direction: on Save the form calls a tool back with tools/call, which the host proxies to the server (gated by the app-callable allowlist):

// inside the form template — call the server's write tool over the bridge
parent.postMessage({ jsonrpc: "2.0", id: 7, method: "tools/call",
  params: { name: "log_feedback",
            arguments: { feature, sentiment, note } } }, "*");
// host replies with the CallToolResult -> re-render the saved list

The server persists the note and returns the updated list, which the form shows — the full form → persist → read-back loop, driven from the Connext side.

Connecting it to Connext

Same as mcp-server-example — this server is its own OAuth provider, so Connext drives standard OAuth 2.1 with dynamic client registration:

1. Expose the server on a public HTTPS URL and run it with PUBLIC_URL set to that URL (every OAuth discovery endpoint is built from it).

2. Register it in Connext (Admin → MCP Servers → Add): URL https://<your-host>/mcp, Transport HTTP, Auth OAuth, client id/secret blank (dynamic registration handles it). Enable Allow UI so the apps render, and mark log_feedback app-callable so the feedback form is allowed to save.

3. Connect as a user — click Connect, sign in on this server's login page, and the agent can call the tools as that user.

Taking it to production

This example keeps everything in memory / in a demo file so it's easy to read. For a real deployment:

• Users: replace DEMO_USERS in auth.py with your real user store + hashed passwords (or delegate to SSO — see the sibling mcp-server-entra-example).

• SQL: point db.py at your real warehouse (Postgres/Snowflake/…) and keep the read-only + single-statement guardrails (add query timeouts + row limits).

• RAG: swap the in-memory BM25 for a real vector store + embeddings (pgvector, etc.) and chunk your documents.

• Charts: the SVG builders scale fine; add a hover/tooltip layer for richer interactivity if your host allows scripts in the MCP App iframe.

• Tokens/HTTPS: persist tokens (or issue signed JWTs) and terminate TLS in front; set PUBLIC_URL to the https:// URL.

The files