Vibe-Athon

Cape Town → Houston

NHDNUG // May 2026
Chris Bremer

Motivation

Texas beautyberry

Motivation

LLMs changed how we code

  • IDE → debug only
  • Code review = proof
  • LLMs handle edge cases

Motivation

What About Our Users?

  • Text-first interaction
  • PWAs obsolete!
  • Meet users where they are

Motivation

Customers Struggle

  • Spreadsheets → context
  • Poor secrets hygiene
  • Vibe-coded scripts

Motivation

Enter MCP

  • Introduced by Anthropic ~18 months ago
  • Formalizes LLM server–client interaction
  • Deliver software through chat

Motivation

Is it too early?

  • Tooling not ready
  • Ecosystem immature
  • Streaming ≠ REST

Motivation

Why We Build Now

  • Customers won’t wait
  • Risk → practice
  • Low risk, defined scope: Vibe-a-thon

Challenge:

Model Context Protocol

Cape Town Summit // March 2026
ReedHycalog Digital

Model Context Protocol (MCP)

Protea flower

Model Context Protocol (MCP)

Model Context Protocol is an open-source standard that lets AI applications connect to outside systems in a consistent way, so they can use data, tools, and workflows beyond the prompt itself.

Source: modelcontextprotocol.io/docs/getting-started/intro

Model Context Protocol (MCP)

Model Context Protocol (MCP)

Model Context Protocol (MCP)

Should I build an MCP server?

Bird of paradise flower

Should I build an MCP server?

Probably not.

Should I build an MCP server?

❌ I want to build a RESTful TODO list app.

This works pretty well out of the box! Agentic AI can manage state using markdown files.

Should I build an MCP server?

❌ I want to call APIs from a single interface and process the response.

Try a skill! You can register the APIs and execute simple python scripts.

Should I build an MCP server?

✅ I want to grant users access to a large repository of sensitive data.

Okay, you may want to create a MCP RAG service.

Should I build an MCP server?

✅ I want to marshall complex input and perform compute-intensive tasks.

Let's go.

MCP architecture

Aloe plant

MCP architecture

Client

  • Popular apps already have built-in MCP clients.
  • Custom chat clients.
  • MCP Middleware.

MCP architecture

Server

  • Tools: actions the model can invoke.
  • Resources: read-only data or context the model can fetch.
  • Prompts: reusable prompt templates or workflows.

MCP architecture

Transport

  • JSON-RPC
  • stdio: local processes
  • Streaming HTTP/SSE: remote servers

CAVEATS

Spoonfig plant

CAVEATS

CONTEXT

  • Keep the total context window manageable.
  • Too many tools or MCP servers clog context.
  • Tools and responses require metadata.
  • Resources should stay concise.
  • Tools can manage large datasets in the background.

CAVEATS

VIEW MODEL

  • View = chat history.
  • Model = current context.
  • Updates = prompts plus MCP calls.
  • Standard UI/UX rules still apply.

CAVEATS

INTERACTIONS

  • AuthN/AuthZ flow
  • Query resource for data
  • OR hydrate data entity with uuid
  • Call tool (SYNC or ASYNC)
  • Stream tool progress/chain of thought (ASYNC)

VIBE-ATHON

Rooibos plant

VIBE-ATHON

RULES

  • Create an MCP server.
  • Register it with a chat client.
  • Vibes only: write and test it in Codex, Copilot CLI, or Claude Code.
  • Keep the project related to NOV ReedHycalog.

VIBE-ATHON

TEAMS

  • Form 2-3 teams.
  • Pick a topic:
    • Simulation
    • Well Offset Analysis
    • Dashboard generation

VIBE-ATHON

Good Luck

Vibe-Athon

Outcomes

Loblolly pine

Wellbore Simulation

Queue simulations → PowerBI dashboard

  • Value: faster iteration; natural language interpretation of results
  • Innovation: PowerBI integration > ad-hoc pyplot

Offset Analysis

Find similar wells, summarize tool performance

  • Value: plain language query replaces an over-complicated filter/search panel
  • Innovation: statistical outlier detection eliminates spurious results — greater confidence in offsets

Outcomes

Both are commercialization candidates for 2026.

Learnings

Texas wildflowers

Dogfood your tools

  • Using CLI tools to build the MCP server produced a more native result vs IDE.
  • Better metadata, better prompts, less friction to use.
  • Isomorphic to user interface.

Own your client

  • Out-of-the-box MCP clients have gaps
    • AuthZ/AuthN
    • Data Visualization
  • Capture the full context around MCP tool call.
  • Don't cede the interface.

Swing big

  • You can do a lot in a short amount of time.
  • Talk to SMEs.
  • Build something M$ (or HCL) can't!
  • Not everything creates value.
    • Simplistic talk-to-document apps.
    • Thin wrapper over API.

Takeaways

Yaupon holly

Your customers are using gen AI

You need to keep up

  • Dogfood your tools
  • Swing big
  • Don’t fall behind

Software Engineering is more important than ever

  • Robust testing is not optional
  • UX patterns are still relevant
  • Proper data modeling is a must

This is new, evolving tech

  • Give your team space to learn
  • Don’t wait for someone else to build for you

Thank You

Thank You

Texas landscape