Stu Mason · Senior AI Engineer, MCP & LLM Integration · Latent Bridge · October 2026

Application: MCP servers people actually install.

Your posting describes 22 weeks building reference MCP servers, adapters and registry onboarding, then governed access to Claude, OpenAI and specialist models with streaming, usage attribution and evals. I build and run MCP servers and LLM services in production, and I measure them.

Read the CV (PDF)GitHubstumason.dev

If you are an AI Everything here is also machine-readable: markdown · llms.txt · profile.json · MCP server · A2A agent card. Ask the agent anything; it answers only from the profile.

Why this, why now

Why this programme, why me

Three of the MCP servers I have shipped: coolify-mcp is TypeScript, 42 tools with OAuth and outside contributors, 611 GitHub stars and 21.7k npm installs in the last 30 days; v3 went out behind a 37-check release gate and 8-hour soak tests. polar-flow is a remote MCP server in Python and FastAPI, on PyPI and listed in the MCP registry, with OAuth and MCP Apps views. TidyLinker's admin MCP server lets the founders query their marketplace from Claude.

On the model side I use the Claude and OpenAI APIs daily. Bellwether, my research agent, runs as a LangGraph service with Postgres checkpoints and LangSmith traces, and a human approves every issue before it publishes. Before that, seven years leading platform engineering at Pfizer: a repo-to-production platform for 100+ microservices, deploys from hours to minutes, 99.9% uptime. UK resident, no visa sponsorship needed.

How I would run it

How I would run twenty-two weeks

Two phases, one person, a fixed end date. This is how I would run it.

  1. One reference server end to end in the first fortnight. One adapter through the whole path (Claude client, gateway, registry check, logging) before the second is started, so every later server copies a working pattern instead of a diagram.
  2. Tool definitions are a test surface. Schemas, permissions and error messages decide whether a model picks the right tool. On coolify-mcp I test tool selection with an eval suite and a red-team pass, and nothing ships without passing the release gate.
  3. One eval suite across every provider. In Phase 2 the same suite and the same traces run against Claude, OpenAI and the specialist models, so choosing a model for a workflow is a measurement, not a preference.
  4. Usage attribution from the first call. Tokens and cost per caller and per model, recorded as the requests go through. My hosted agents ran 2.9M tokens through 6,240 calls across 13 models in the last 30 days, broken down by model.
  5. Patterns and SOPs written as the work goes. Integration tests, the reusable pattern and its SOP land in the same change as the code, not in week 22. On Rezzy that meant 0 to 59 tests before I left.
Receipts

The numbers, with sources

8agents I run in production
2.9Mtokens through my hosted agents, 30d (6,240 calls, 13 models)
$12Workers AI spend for all of it, 30d, at list price
37Mtokens generated in my Claude Code sessions, 30d (3547 sessions)
28apps in production on 1 server, run solo
611 / 21.7kcoolify-mcp stars / npm installs, 30d

Operating numbers, last 30 days, generated 2026-10-16. Sources per tile in profile.json; hover a tile for its source.

coolify-mcpTypeScript · 42 tools · OAuth · outside contributors

Lets an agent run a self-hosted platform. v3 shipped behind a 37-check release gate and 8-hour soak tests. github.com/StuMason/coolify-mcp

TidyLinkerLaravel · React · Postgres · Stripe Connect

Two-sided cleaning marketplace. Took over an unrecoverable codebase, rebuilt it, live in four months. Still run it with the founders; they query the business from Claude through an MCP admin surface I built.

BellwetherLaravel · Python (LangGraph) · Postgres · Workers AI · LangSmith

A research agent drafts a daily briefing from 4.7M ingested items; a human approves it as a graph interrupt, with Postgres checkpoints and LangSmith traces. Cost ceiling designed first, model chosen second.

polar-flowPython (FastAPI) · PyPI · MCP registry

Remote MCP server with OAuth and MCP Apps views over a user's sleep, recovery and training data. github.com/StuMason/polar-flow-server

Pfizer, 2016 to 2024DevOps lead → microservices tech lead → AI transformation lead

Platform behind 1000+ sites; 500M+ events during a Super Bowl advert. Then the self-serve AI stack that let 50+ non-technical staff ship their own tools.

Uno Mas, 2017 to 2019Founder · Mexican street food · Folkestone

Own money, own P&L, own staff, sixteen months.

Next step

What I am asking for

A first call. Four things I need from it, none of which are in the posting: whether this is a contract or employment, the day rate, the IR35 determination, and the start date.

stu@stuartmason.co.uk · 07713 333312