Decision. Build a read-only database query server (SQLite or Postgres). It’s the only single idea that naturally exercises all three core primitives —
tools(run SQL),resources(expose table schemas asschema://...URIs),prompts(canned “analyze this table” templates) — needs zero external auth (ship a seeded.db), and lands inside 2-3h for a senior audience that already thinks in schemas and queries [1][2]. Backup pick if you want a sampling “wow” moment: a Hacker News digest server that uses MCP sampling to make the client’s LLM summarize stories the server fetched — a server with no LLM key of its own acting like an agent [3][4].
Why these constraints matter
A live session for Java/.NET seniors using the official TypeScript SDK ⭐ 12k (Jun 2026) [5] needs an idea that (a) goes past the canonical “weather tool” everyone has seen [6], (b) shows resources and prompts, not just tools — most reference servers only demo tools, and the official Everything server is the only one that bundles all three [7], and (c) avoids OAuth dances and infra that eat the clock. The SDK’s registerTool() / registerResource() / registerPrompt() high-level API auto-handles capability negotiation and Zod input validation, so the primitives themselves are cheap to wire — the time goes into the domain behind them [8][9].
Auth is the usual time-sink. Free, keyless public APIs (Open-Meteo, Hacker News Firebase/Algolia, Wikipedia) remove that entirely [4][10]. Anything behind OAuth (Spotify, Gmail, GitHub write) costs you a token flow on stage — possible via the SDK’s URL elicitation, but a poor first-session choice [11].
Shortlist scored
Scores 1-5 (5 = best/easiest). “Primitives” = which of tools (T) / resources (R) / prompts (P) / sampling (S) / elicitation (E) the idea naturally exercises.
| Idea | Build-in-time | Wow | Utility | Auth complexity (5=none) | Primitives | Notes |
|---|---|---|---|---|---|---|
| DB query server (SQLite/Postgres) [1] | 5 | 4 | 5 | 5 (ship seeded .db) |
T·R·P | Schema as resources, SQL tool read-only, “analyze table” prompts. Senior audience’s home turf. |
| HN digest + sampling [3] [4] | 4 | 5 | 3 | 5 (keyless HN API) | T·S | Server asks client LLM to summarize fetched stories — the “servers can be agents” reveal. |
| Interactive incident form (elicitation) [12] | 4 | 5 | 4 | 5 (local) | T·E | Tool pauses mid-call to elicitInput() a structured form — multi-turn UX, very demo-friendly. |
| Markdown/PDF → structured doc [13] | 4 | 3 | 4 | 5 (local files) | T·R | Convert + expose result as a resource. Tangible output, modest wow. |
| Git repo insights (local repo) [14] | 3 | 4 | 5 | 5 (local git) | T·R·P | Commit/diff tools + repo file resources + “review this diff” prompt. ⚠ git edge cases eat time. |
| Open-Meteo weather [10] [6] | 5 | 2 | 3 | 5 (keyless) | T | The reference build — fast, but everyone’s seen it. Use only as a 10-min warmup. |
| Browser automation (Playwright) [15] | 2 | 5 | 4 | 5 (local) | T | Highest wow, but reimplementing it is out of scope; Microsoft’s server already exists. ⚠ flaky live. |
GitHub stars for repos referenced above: mcp-sqlite ⭐ 107 (Jun 2026) [1]; official servers ⭐ 87k (Jun 2026) [7]; playwright-mcp ⭐ 33k (Jun 2026) [15]; Open-Meteo-MCP-Server ⭐ 0 (Jun 2026) — reference only, not a popular project [10].
Top pick — DB query server, session shape
A seeded SQLite database (e.g. startup-funding or Titanic sample data — both are common in existing servers) gives you a self-contained demo with no network or keys [1]. The build maps cleanly onto a 2-3h arc:
- Tools (~45 min).
list_tables,describe_table,query— gatequerytoSELECT/PRAGMA/EXPLAIN/WITHonly, the standard read-only safety pattern, which also seeds a good “AI safety in tools” discussion for a senior crowd [1]. Zod schemas give the .NET/Java audience the familiar “DTO validation” mental model [8]. - Resources (~30 min). Expose each table’s schema at a
schema://sqlite/<table>URI; resources are read-only context the client pulls in, distinct from tool actions [2]. This is the part most tutorials skip — showing it differentiates the session. - Prompts (~20 min). Register an
analyze-tableprompt template the user can pick from a menu, prefilled with the schema — demonstrates the third primitive that almost no demo covers [7]. - Wire to a client (~20 min).
StdioServerTransportinto Claude Desktop or any MCP client; ask “which 3 sectors raised the most in 2025?” and watch it discover schema → write SQL → answer [6]. - Buffer/stretch. Swap SQLite for a Postgres connection string to show the same server against real infra, or add an
elicitInput()confirmation before any expensive query.
The .NET/Java audience already has the database mental model, so cognitive load goes entirely into MCP mechanics, not the domain — exactly what you want for a first build [8].
If you want one “wow” beat — add sampling
If the room is advanced and you finish the DB server early, the standout reveal is sampling: the server asks the client’s LLM to do work, flipping the usual flow [3]. Concretely: a hn_digest tool fetches top Hacker News stories (keyless hacker-news.firebaseio.com / Algolia endpoints) [4], then issues a createMessage sampling request so the client LLM summarizes them — the server itself holds no API key and runs no inference, yet behaves like an agent [3][16]. The SDK ships runnable sampling and elicitation examples (elicitationFormExample.ts, elicitationUrlExample.ts) you can crib from live [11]. ⚠ Client support for sampling is uneven — verify your demo client (Claude Desktop) honors it before relying on it on stage [16].
What to avoid for a first live session
- OAuth-backed APIs (Spotify, Gmail, GitHub write) — token flow burns stage time even with URL elicitation [11].
- Browser automation from scratch — highest wow but Microsoft’s playwright-mcp ⭐ 33k (Jun 2026) already does it; reimplementing is out of scope and flaky live [15].
- Bare weather/echo tools — fine as a 10-minute warmup, but they don’t show resources or prompts and the audience has seen them [6].