AI engineering studio

AI agents, integrations & apps that actually ship.

Cybrix.AI designs, builds and runs production AI software: agents that use your tools, MCP servers and API integrations, ChatGPT plugins, and the self-hosted infrastructure underneath. One engineer, from first sketch to production.

26 ChatGPT plugins built OpenAI + Claude Custom GPTs retire in 76 days
Terminal illustration: David Cahill of Cybrix.AI LLC offers AI agents, MCP integrations, ChatGPT plugins, development, automation and infrastructure; 26 plugins listed in plugins.json.
Model Context ProtocolOpenAI APIClaude APIClaude Agent SDKMCP Apps widgetsOAuth 2.1 / OIDCPython · FastAPITypeScript · Next.jsPostgreSQLDocker ComposeApache · CaddysystemdCloudflareGitHub Actions
01 / services

From a single integration to the whole stack.

Bring a problem, a half-built prototype or a system that needs an AI layer. I'll design it, build it, deploy it and keep it running.

01agents

AI agents & harnesses

Agents that do real work with your tools and data, plus the harness around them: orchestration, guardrails, logging, retries and human approval steps.

  • Claude & OpenAI agent SDKs
  • Tool use, retrieval, memory
  • Evals, tracing, cost controls
02integrations

MCP servers & integrations

Give ChatGPT, Claude or your own agent safe, typed access to your systems through MCP servers, REST APIs, webhooks and proper OAuth sign-in.

  • Remote MCP servers (Python, TypeScript)
  • OAuth 2.1 · PKCE · JWT
  • OpenAPI → MCP conversion
03chatgpt

ChatGPT plugins & apps

Plugins that live inside ChatGPT with interactive widgets: diagrams, slide decks, audio, quizzes, games. I've built 26 of them.

  • Inline MCP Apps widgets
  • Directory review readiness
  • Custom GPT → plugin migration
04build

Custom development & automation

Full-stack web apps, internal tools, document-generation pipelines and scheduled jobs that remove the manual steps from a workflow.

  • Python · FastAPI · TypeScript · Next.js
  • PPTX, DOCX, PDF, XLSX, ICS output
  • Data pipelines & reports
05infra

Infrastructure & DevOps

Self-hosted and observable. Linux servers, reverse proxies, DNS and TLS, systemd services, containers, CI/CD and uptime monitoring.

  • Ubuntu · Apache · Caddy · systemd
  • Docker Compose · GitHub Actions
  • DNS, TLS, Cloudflare, status pages
→deadline

Custom GPTs retire December 11, 2026

GPT Actions don't carry over to plugins. I've already moved my own catalog; I can move yours, with its tools rebuilt as an MCP server.

Plan your migration
02 / selected work

Real systems, built and running.

No mock-ups: these are Cybrix.AI's own products and the infrastructure behind them. The same patterns go into client work.

26 plugins · one shared kit3 coming soon  23 testing  
case / 01 · chatgpt plugins

One MCP kit, 26 ChatGPT plugins

When OpenAI announced that custom GPTs would retire, I rebuilt the whole Cybrix.AI catalog as ChatGPT plugins. A shared Python kit turns existing FastAPI backends into MCP tools with an inline widget, so each backend serves the old GPT and the new plugin side by side until the cutoff.

  • Widgets in the chat: molecule and network diagrams, slide carousels, sheet music with an MP3 player, quizzes, a blackjack table
  • One script provisions a new app: DNS record, Apache vhost, TLS certificate, systemd service
  • Every tool parameter documented and checked by an automated review-readiness inspector (case 03)
  • Python
  • FastAPI
  • MCP
  • MCP Apps
  • Apache
  • systemd
GET /.well-known/oauth-authorization-server
{
  "issuer": "https://auth.cybrix.ai",
  "registration_endpoint": ".../register",
  "grant_types_supported": ["authorization_code", "refresh_token"],
  "code_challenge_methods_supported": ["S256"]
}
case / 02 · auth

OAuth 2.1 sign-in for plugins

Plugins that keep per-user data (time tracking, saved prompts, a game character) need real accounts, so I built a standards-based authorization server instead of bolting on API keys.

  • Dynamic client registration and PKCE, the flow ChatGPT expects
  • RS256 tokens scoped to each plugin's origin, rotating refresh tokens, public JWKS
  • Google sign-in that stores only a hash of the account ID
  • OAuth 2.1
  • OIDC
  • JWT
  • MCP SDK
› inspect_mcp_server https://chemistry.cybrix.ai/mcp
  ✓ handshake     streamable HTTP
  ✓ tools         valid names, titles
  ✓ annotations   readOnly · destructive · openWorld
  ✓ schemas       every parameter described
  ✓ widgets       resources readable
  readiness_score 100
case / 03 · developer tooling

MCP & Plugin Builder

A plugin that builds plugins. It connects to a live MCP server and checks it against the directory rules, validates plugin packages, and scaffolds new servers in Python or TypeScript, including converting a GPT Action's OpenAPI spec into MCP tools.

  • Readiness score and a downloadable scorecard for any server
  • Scaffolds for both the Python and TypeScript MCP SDKs
  • MCP
  • OpenAPI
  • Python
  • TypeScript
case / 04 · domain tools

Chemistry Studio

Publication-quality molecules, curved-arrow mechanisms and energy profiles drawn right in the chat. It balances any equation including redox half-reactions, runs limiting-reagent, yield, pH and buffer calculations, and pulls PubChem data with GHS safety information.

  • RDKit
  • PubChem API
  • SVG rendering
  • FastAPI
case / 05 · multiplayer

MMOGPT

A persistent multiplayer RPG inside ChatGPT. The rebuild moved all game state out of the model and onto the server, so every player explores the same procedurally generated world, with inventory, encounters, daily events and rendered maps. The original world was imported intact.

  • Game server
  • SQLite
  • Pillow
  • OAuth
Plus plugins for SOPs, network diagrams, slide decks, music, SEO audits, SEC filings, NIST 800-53 and more. Browse all 26 plugins
03 / how I work

Short loops and working software, early.

You talk to the engineer who writes the code. Decisions are made on something you can click, not on slides.

  1. step 01

    Scope

    A call or an email thread to pin down the problem, the systems involved and what "done" looks like.

    → written plan & quote
  2. step 02

    Prototype

    A working version early, against your real data or a safe copy, so the hard questions surface first.

    → demo you can use
  3. step 03

    Build & harden

    Auth, error handling, tests, logging and monitoring: the parts that decide whether an AI feature survives real users.

    → tested release
  4. step 04

    Ship & hand over

    Deployed where it should live, on your cloud or servers or hosted for you, with docs and runbooks.

    → production + runbook
04 / toolbox

Tools I build with daily.

Chosen for boring reliability. If your team already runs something else, I'll work in it.

AI & agents

  • OpenAI API
  • Claude API
  • Claude Agent SDK
  • Model Context Protocol
  • MCP Apps widgets

Backend & web

  • Python · FastAPI
  • Pydantic
  • TypeScript · Node
  • Next.js · React
  • Django

Data & auth

  • PostgreSQL
  • SQLite
  • Redis
  • OAuth 2.1 · OIDC
  • JWT · JWKS

Infrastructure

  • Ubuntu Linux
  • Apache · Caddy
  • systemd
  • Docker Compose
  • Cloudflare · GitHub Actions
05 / about

A studio of one, by design.

Cybrix.AI LLC is run by David Cahill. You work directly with the engineer who designs, builds and deploys your project: no account managers and no hand-offs.

I've been building on OpenAI's platform since custom GPTs launched in 2023. Today I run a catalog of 26 ChatGPT plugins, each an MCP server on its own subdomain, on infrastructure I set up and operate myself.

That same stack of agents, MCP, OAuth, document generation and Linux operations is what I bring to client work.

Quick facts about Cybrix.AI
# about.yml
company:  Cybrix.AI LLC
engineer: David Cahill
focus:    [agents, mcp, integrations,
           chatgpt-plugins, infrastructure]
plugins:  26   # from plugins.json
since:    2023 # first custom GPTs
models:   [openai, claude]
status:   status.cybrix.ai
contact:  info@cybrix.ai
Contact

Have something to build? Let's talk.

Email is the quickest way to reach me. A few lines is plenty to start. I'll reply with questions or a suggested next step.

Useful to include

  1. What you want the AI or software to do, and for whom
  2. The systems it has to talk to (APIs, databases, SaaS tools)
  3. Where it should run: your cloud, your servers, or hosted
  4. Any deadline, such as the December 11 GPT cutoff