ServicesMCP server development
MCP server development to connect internal APIs to AI assistants
MCP, the Model Context Protocol, is a common interface through which an AI assistant calls tools and reads context from other systems. We build MCP servers that give an assistant access to the internal APIs you approve, inside your environment and with your team. The protocol standardises the connection; what the assistant may do through it, we design with you.
What happens when the assistant calls a tool
- 1
The assistant sends a request that matches the tool's defined input.
- 2
Access is checked against the permissions agreed for that tool.
- 3
Reads return data. Writes wait until a person approves them.
- 4
The call is logged without credentials, and failures return a clear error.
An example with support tickets
Suppose your support team wants its assistant to look up a customer's open tickets and draft a status update. An MCP server could offer two tools. One reads ticket details through your helpdesk API with the access rights of the person asking. The other prepares the update and posts nothing until that person has approved the text.
Starting with one tool
First we check whether you need an MCP server at all. Your assistant may already have a connector for the system, or an ordinary API integration may be simpler. When a server is the right choice, we start with one bounded tool and the client your team actually uses.
Every tool gets a schema that states what it accepts, what it returns and what each error means, so the model can use it correctly and a reviewer can see what happened.
Access, identity and approvals
We begin with read-only tools and the narrowest scopes the task allows. Where the underlying system supports it, a call runs as the user who asked, so your existing access rules apply. Any tool that changes data first asks a person for explicit approval.
Content a tool returns, such as a customer's comment on a ticket, can contain instructions aimed at the model. We test each tool with content like that. Permission checks and approvals are enforced in the server and the systems behind it, so they do not depend on the model ignoring those instructions.
- Write access added one tool at a time, after the read-only version works
- Tokens and secrets kept out of logs and tool output
Failures, retries and tests
Internal APIs time out, hit rate limits and sometimes fail halfway. Each tool gets timeouts and retry rules that respect the API behind it, and write operations are idempotent, so a repeated call does not post the same update twice.
Tests cover allowed and denied inputs, partial failures, and the protocol and client versions you run. Clients implement different parts of MCP, so we confirm what works with yours.
Where it runs and what you keep
Whether the server runs locally next to the client or as a remote service in your environment is decided after we have reviewed access and security with your team.
- Source code and configuration in your repositories
- Tests your CI can run on every change
- Runbooks for adding a tool and revoking access
- An agreed owner on your side for each tool
Questions
What happens in the first call?
Thirty minutes with one of our engineers, free of charge, about the systems you want to connect and the assistant your team uses. We need no access to any of them. Afterwards we propose a scope and an estimate.
Does MCP make an assistant secure?
Not on its own. MCP defines how an assistant and a tool talk to each other. Safety depends on the permissions, approvals, logging and tests around each tool, and that is the part we design with you.
Will the server work with every AI client?
No. Clients support different parts of the protocol and change between releases. We test against the clients you name and tell you where support is limited.
Do you sell ready-made MCP servers?
No. We build servers for your own APIs and systems as a project, and the code goes to your team at handover.
Related services
AI automation
Business processes automated step by step with rules, models and agents, tested on past cases and reported as cost per case.
AI chatbots
Chatbots and assistants for customer service and internal support that answer from your data, act through approved tools and hand over to people.
AI consulting
AI strategy, use case assessment and readiness checks from engineers who also build and measure the first system.
AI engineers for your team
Experienced AI engineers who join your team part-time or for longer and build AI features, agents and platforms with you.
Enterprise AI
Assistants, agents and private AI built into your systems, with your permissions, audit logs and a platform team that can run them.
RAG development
Assistants that answer from your internal documents, with permissions, citations and tests on your own questions.
AI engineering and architecture
Assistants, voice interfaces and agents in the products you sell, with the serving and retrieval underneath.
Custom model training
A small open-weight model trained for one task and compared with your current model on cost and accuracy.
From the blog
MCP tool design
Why an MCP server should usually offer a few tools shaped around business actions instead of mirroring every API endpoint, with a worked example and a tool contract.
MCP server permissions
How to connect an assistant to internal systems without giving every MCP tool broad access: scopes, checks inside each tool and separate downstream identities.
MCP server test plan
A test plan for an MCP server before users connect: contract tests, negative authorisation tests, failure tests and state-change tests, with checks that run in CI.
Talk to an engineer
Bring one internal system you would like an assistant to use. The call takes 30 minutes and needs no access to your systems or your data.
Booking runs on Google Calendar.
Prefer to write first? Leave your email and a few lines, and we reply with times.
Look for the draft in your email program
Send it from there, and an engineer replies in person with times for a 30-minute call. If no draft has opened, write to us at miki@cloudsail.com.