Fractional CTO · MCP-Server Playbook

I ship your product as a remote MCP server, and stay on as your fractional CTO after.

You need a technical adult in the room. Not full-time, not for the next ten years, just until you know what you are doing. Two products, one operator, natural funnel.

See the pattern before you buy it. The MCP-Server Playbook ships the exact architecture already running at github.com/tounsils/ask-me-mcp, a public MIT-licensed reference implementation. Install it in Claude Desktop or Claude Code and interact with the pattern during our discovery call: six typed tools, a coordinator, rails and confidence scoring, external grounding, an evaluation corpus and a certification gate. Every element of the pattern is visible in the source.

Live endpoint: https://ask-me-mcp-xi.vercel.app/api/mcp

Product 1: MCP-Server Product Playbook

Fixed fee 6 weeks · quoted on the call Ships or 50% refund

Your product ships as a remote MCP server in Anthropic's connector directory, ChatGPT's Apps SDK, and via xAI's tool-calling API. One MCP server. Three storefronts. Same code.

What you get in 6 weeks

  1. Typed tool contract. Four to eight JSON-schema-strict MCP tools. No free-form params, no prompt-engineering surface area.
  2. Coordinator and specialist agent architecture. The coordinator routes, elicitation agents gather, supervisor agents review before the response goes out.
  3. Signal-extraction pipeline. Every conversation compiles to a typed signal vector. Structured, versioned, inspectable.
  4. Versioned reasoning specification. Grounded in your domain science, not the model's guesses. Whatever framework governs your field, we ship it as spec.
  5. Rails and confidence-per-answer discipline. Steering blocked in code. Every response ships with a confidence level and its evidence.
  6. Grounding architecture. The harness queries your data sources, external APIs, and reference taxonomies. The model never invents.
  7. Evaluation corpus, eval runner, per-model-pairing certification. You author the corpus with my guidance. I ship the runner and the gate. Pass or nothing ships.
  8. Remote MCP server deployment. Production endpoint, auth, observability, and the connector-directory listing paperwork.
  9. One passing eval on your production tool contract by week 6.

Week by week

WeekDeliverableGate
1Product definition, candidate tool list, input/output schemas, architecture mapYou validate the target user flow and the tool contract. No tool contract without schema review.
2Grounding sources identified, confidence and refusal policy in code, first versioned reasoning specDomain science is explicitly encoded, not left implicit. Critical edge cases documented.
3Coordinator routes, specialist workers, handoff behavior, first evaluator passTool routing is inspectable and deterministic. The system can explain which path it used.
4Evaluation corpus, eval runner, documented failure cases and known gapsEvery core tool has a smoke eval and a failure case. You confirm behavior matches product intent.
5Auth, deployment, observability, production logging, final fallback logicA deployable remote MCP endpoint exists. Critical failure modes are handled and visible.
6End-to-end smoke tests, final eval pass, connector-directory launch checklist, handoff notesOne passing eval on the final production tool contract, or the refund trigger fires.

Not in scope

  • Your domain implementation. That is your product, the tools call it.
  • UI beyond a debug console.
  • Marketing, growth, or go-to-market. That is what the retainer is for.
  • Fundraise deck work.

Refund trigger

  • No live tool contract on the target remote MCP setup.
  • No working eval runner or validation gate.
  • No production-ready grounding architecture.
  • No deployable path to the connector directory.

Failure to reach the shipping state by end of week 6 is failed delivery, not a soft "almost shipped."

Who this is for

Pre-seed to Series A founders who have already decided that going deep on one AI-assistant platform beats building yet another web app. You value defensibility from external certification, not just "we use Claude."

Who this is not for

  • Teams that want a chatbot with a retrieval tab.
  • Teams without a real domain to encode. The harness is only defensible if you have field science underneath.
  • Teams that cannot commit an engineer or PM to the sprint. This is not turnkey.

Cap: one active Playbook engagement at a time. When at capacity, waitlist only.

The architecture you are buying

MCP-server harness architecture: typed tool contract, coordinator and specialist agents, signal vector, versioned reasoning spec, rails and confidence, external grounding, evaluation corpus and certification gate.
The Harness pattern. Every layer is delivered in the 6-week Playbook and visible in the ask-me-mcp source.

Product 2: Fractional CTO Retainer

Monthly retainer quoted on the call Cancel anytime

For founders who complete the Playbook and want to keep going, or who need the architecture-adult role without the MCP-shipping wedge.

What you get

Explicitly not in scope

Cap: two active retainer clients at a time.

Terms

ProductBasisTermCap
MCP-Server Product PlaybookFixed fee, quoted once6 weeks1 active at a time
Fractional CTO RetainerMonthly, cancel anytimeMonth to month2 active at a time
Investor or all-hands sessionPer sessionAdd-onBy arrangement

On price. Both products are fixed and quoted, not hourly and not estimated. I do not publish the numbers, because the right scope is worth establishing before the figure is, and a rate card invites comparison against work that is not the same work. Ask on the first call, or by email, and you get a straight answer in one line with no qualifying dance.

Combined ceiling: one Playbook plus one Retainer, or two Retainers, or two back-to-back Playbooks. Never two concurrent Playbooks. I would rather turn work away than deliver two sprints badly.

The proof: current engagements

NXT Robotics: RobotogoAI operator platform

Senior Software Engineer, fractional 60%, 2025 to present.

Build RobotogoAI, the operator SaaS console and agent layer that turns raw device telemetry (video, GPS, battery, BLE beacons) into actionable ops decisions for physical-security and field-equipment fleets: stadium venues including Snapdragon Stadium and Petco Park, EV-charging networks, drone operators, and a 300+ site parking pilot.

Own five sub-modules across the platform: a React 19 operator console (Vite, Ant Design, Redux Toolkit, Saga), a Laravel 12 / PHP 8.3 API with Passport auth, a Python and AWS SAM per-tenant deployment pipeline, a Node 20 executive cockpit on Lambda with Aurora Serverless v2 and Cognito, and a dual-audience knowledge repository read by both humans and coding agents.

Recent ships: an end-to-end alerts UI with severity taxonomy and deep-linking, an iOS video playback fix, BLE beacon CRUD, a battery-low threshold service, a telemetry retention pipeline, a Cognito SRP auth migration, and an admin team-management route with invite flow.

AIMIA: AI career and AI-literacy assessment

VP of Engineering, advisory, 2025 to present.

Own the entire engineering surface for a career-guidance harness serving students, young professionals, and adult learners. Institutional partners include Irvine Valley College, MIT, Palomar College, and Mission Edge.

Architected a multi-agent stack (six specialized agents plus a supervisor synthesizer), bidirectional voice over WebSocket with a fallback provider, multilingual output in English, Traditional Chinese, and Spanish, application-layer AES-256-GCM field encryption with an HMAC-SHA256 blind index for email dedupe, and a length-bias anti-cheat detector. Currently leading a container-platform migration between two major clouds.

The reusable-patterns library

Twelve patterns worth stealing, built across the engagements above. Retainer clients get access plus a walkthrough on the ones relevant to their stack. Selected examples:

Full list on request.

How to book

First 30 minutes are free. A fit check, and an honest answer on whether either product actually solves your problem. If neither does, I will say so.

Email me to book Connect on LinkedIn