S-07 — Service · AI agents
Custom MCP servers that give AI safe access to your business
The Model Context Protocol (MCP) is the standard way to connect AI assistants like Claude to real tools and data. A well-built MCP server lets your team ask Claude about customers, orders or projects, and take approved actions, without copying between tabs. I build them with authentication, permissions and audit logs from the start.
Diagnosis
Sound familiar?
- 01
AI can't see your data
The assistant is smart, but it knows nothing about your customers, your pipeline or your documents.
- 02
Your team pastes context all day
Exporting, pasting and re-explaining the same information in every chat.
- 03
Off-the-shelf connectors don't fit
Your CRM fields, internal database and custom tools aren't covered by generic integrations.
- 04
Security is a blocker
Nobody wants an AI with admin access to production. Rightly.
Scope
What I build
S-07.1
Custom MCP servers
Tools and resources that expose exactly the data and actions your team needs, and nothing more.
S-07.2
Remote MCP with OAuth
Hosted servers your whole team can connect to, with sign-in and per-user permissions.
S-07.3
Read-first design
Read tools by default. Write actions are scoped, confirmed and logged.
S-07.4
MCP for your SaaS
A public MCP server, so your customers can use your product from Claude and other AI clients.
S-07.5
Internal tool bridges
Wrap legacy databases, spreadsheets or internal APIs so AI assistants can use them safely.
S-07.6
Docs & rollout
Setup guides for Claude Desktop, Claude Code and other clients, plus a walkthrough for your team.
Method
How the engagement runs
MCP design is permissions design. I start by mapping which people need which data and actions, then shape the tools around that. Small, well-described tools with clear inputs work far better for AI, and are far safer, than one giant do-anything endpoint.
- 01
Survey
I map how work actually moves today: every tool, handoff, spreadsheet and inbox. You get the map in writing before anything is built.
- 02
Blueprint
A scoped plan: what gets automated, what stays human, what it connects to and what it will cost to run. You sign off before I build.
- 03
Build
Delivered in stages your team can test with real data, with error handling and logging from day one rather than bolted on later.
- 04
Handover
Documentation and a recorded walkthrough, so your team can run it and change it. Your accounts, your data, your system.
Questions
Questions people ask
What is an MCP server, in plain English?
A small service that tells an AI assistant which tools it can use, like "look up a customer" or "create a task", and then carries out those requests against your systems within the permissions you set.
Which AI apps can use it?
Claude (desktop, web and Claude Code) supports MCP natively, and so do a growing number of other AI clients and agent frameworks. One server can serve all of them.
Is it safe to connect AI to our systems?
It can be, if the server is designed for it: least-privilege access, per-user sign-in, confirmation on write actions and audit logs. That's the default in everything I build.
Should the MCP server run locally or remotely?
Local servers are quick for personal or developer use. Remote servers with OAuth are better for teams: one deployment, central permissions and no setup on each laptop.
Contact