SystemsSheet 04 of 08

MCP servers that give every AI tool your company's skills

The Model Context Protocol (MCP) is the open standard AI tools use to connect to data and actions. We build and host MCP servers that carry your approved skills, your data and your website, so Claude, Cursor, Copilot, ChatGPT and your own agents all work from the same source.

Serves: Skills, knowledge, website content and actions
Clients: Any MCP-compatible AI tool
Transports: Streamable HTTP, SSE, stdio

A governed MCP hub for skills, data and your website

Switch the environment to see the services it runs on. Play the walkthrough, or pick a step.

Deterministic codeAI stepHuman reviewYour existing systemAI assist, approved by a person

A governed MCP hub for skills, data and your website

One governed endpoint between every AI tool and your skills, data and website, with authentication and an audit trail.

Our own hub runs this way: Docker on Render, bearer tokens, host allowlist.

See it run

What actually crosses the wire when an AI tool uses your MCP server to search your website and load a skill.

MCP sessionJSON-RPC between an AI client and your serverIllustrative example

AI clientClaude Desktop

MCP gatewayauth and allowlist

Your MCP servertools and skills

Company websitesearch API

initialize

token valid, host allowed

initialize

capabilities: tools, resources, prompts

tools/list

4 tools, 6 skills

tools/call search_site

GET /api/search?q=private+LLM

3 results

result: 3 pages with links

resources/read skill://data-engineering

SKILL.md, version 1.4

Request, message 7

{"jsonrpc": "2.0", "id": 7, "method": "tools/call",
 "params": {"name": "search_site",
   "arguments": {"query": "private LLM for data mapping"}}}

Response, message 10

{"jsonrpc": "2.0", "id": 7, "result": {
 "content": [{"type": "resource_link",
   "name": "Data engineering automation",
   "uri": "https://yourssherpa.com/data-engineering.html"}],
 "isError": false}}

Component view

The same system as an exploded 3D drawing. Each plate is one step, and each block is one component in one of four materials.

Drag to rotate. Select a layer to inspect it.

Parts list

Numbered bottom to top

What agents need to reach: your website and documentation, product or service data, internal APIs and databases.
- Company website Your existing system
- Internal APIs Your existing system
- Databases Your existing system
- SaaS tools Your existing system
Typical toolingCMS, REST and GraphQL APIs, SQL databases, SharePoint, Google Drive
Each connector turns a system into MCP tools and resources with typed inputs. Read tools and write tools are separate, so write access can be granted on its own.
- Website reader Deterministic code
- API adapters Deterministic code
- Query tools Deterministic code
Typical toolingTypeScript and Python MCP SDKs, Zod and Pydantic schemas
Skills are written instructions and playbooks tuned to your work, such as how your team builds a Databricks workflow or answers a policy question. They live in Git, are tested against example tasks, and are reviewed by the team that owns them.
- SKILL.md packs Deterministic code
- Skill tests AI step
- Owner review Human review
Typical toolingGit, SKILL.md packs, OKF bundles for reference knowledge
One server exposes the catalog as MCP tools, resources and prompts over Streamable HTTP or SSE, and over stdio for local use.
- Tool registry Deterministic code
- Resources and prompts Deterministic code
- HTTP and SSE transport Deterministic code
Typical tooling@modelcontextprotocol/sdk, Express, FastMCP, Docker
Every request is authenticated. Tokens are scoped per team, hosts are allowlisted, and calls are rate-limited and logged.
- Token or OAuth check Deterministic code
- Host allowlist Deterministic code
- Rate limits and logs Deterministic code
Typical toolingBearer tokens with constant-time checks, OAuth 2.1, TLS at the reverse proxy
Any MCP-compatible client connects to one URL and gets the same catalog on every device.
- Claude Your existing system
- Cursor and VS Code Your existing system
- ChatGPT and others Your existing system
- Your own agents Your existing system
Typical toolingClaude Desktop and Claude Code, Cursor, VS Code Copilot, ChatGPT connectors, custom agents

Deterministic codeAI stepHuman reviewYour existing system

Sheet04 of 08

Layers6

SystemMCP servers and skills

Drawn byA. Adhikari

IssuedSeptember 2026

ScaleNot to scale

Three ways MCP helps

MCP works in both directions: it lets AI tools reach you, and it lets your agents reach everything else.

Your company, readable by AI tools

We expose your website, documentation, catalogue and actions such as quotes or bookings as an MCP server. Any agentic tool your customers or partners use can then read your content and act through tools you control, instead of scraping pages.

Your agents, connected to any website

Internal agents reach outside services and websites through MCP connections, with the same authentication, logging and review as internal tools.

Your team, working from the same skills

Engineers and analysts use the same approved skills in every AI tool, so results no longer depend on who wrote the prompt.

Skills tuned for your use case

A skill is a reviewed playbook the AI tool loads when the task calls for it. These run today in our own hub, followed by two examples of what we build for clients.

data-engineeringStandards and review patterns for pipeline work.
databricks-secretsSafe handling of secrets and configuration in Databricks.
databricks-workflow-builderComposing Databricks job task graphs and workflow definitions.
delta-table-operationsDelta table lifecycle, migrations and reliability checks.
brand-and-toneYour voice and formatting rules for anything customer-facing. Client example
policy-answersApproved answers to policy questions, backed by an OKF bundle. Client example

Where it runs

The drawing stays the same. The services change with the environment you already run, and the feasibility analysis picks the fit.

Environment

What we use

When it fits

Container platform

Built before

What we useDocker on Render, Fly.io, AWS ECS, Azure Container Apps or Cloud Run

When it fitsOur own hub runs on Render with Docker.

Your private network

What we useBehind your VPN, with OAuth through your identity provider

When it fitsInternal-only skills and data.

Databricks

What we useDatabricks Apps hosting, Unity Catalog functions as tools

When it fitsData teams already working in Databricks.

Local

Built before

What we usestdio transport on each laptop

When it fitsDevelopment and testing.

Feasibility first

Before anything is built, we check whether this system is worth building for you, and where it should run.

Which AI tools your teams and customers use, and which of them support MCP.
Which skills, playbooks and knowledge are worth centralising first.
Which systems and website actions to expose, with read or write access for each.
How people authenticate, and how access maps to your identity provider.
Hosting, logging and review requirements.

What you receiveA catalogue plan listing the first skills and tools, an access model, the hosting design, and a go or no-go recommendation.

Built before

Work delivered by Ashish Adhikari, who leads engineering at YoursSherpa.

Central AI agent hub: a hosted MCP server that serves versioned SKILL.md packs to Cursor, Claude and other MCP clients from one endpoint, with bearer authentication, host allowlisting and Docker deployment.Open source
Graphify knowledge graphs, queryable by AI coding assistants through an MCP server.Open source

Start with a feasibility call

Tell us about one process or data problem. In the first call we will say which parts we would automate with code, which need an agent, and which we would leave alone.

Send a short note through the contact form and we will set up the call.

What helps us prepare

The process or system you have in mind, and who works on it today.
Where the data lives: cloud, platform and main tools.
Security or hosting rules we need to work within.