MCP server development

MCP servers that let AI agents work safely inside your software

The Model Context Protocol lets AI assistants like Claude and ChatGPT use your software directly: look things up, create records, run workflows. We run two MCP servers in production, each with its own OAuth server, so we know how to make that useful without handing an AI the keys to everything.

  • OAuth 2.1 with PKCE and dynamic client registration
  • Tools scoped to the user's role, plan and workspace
  • Human approval for anything that sends, spends or deletes

How an MCP connection works

  1. AI assistant

    Claude, ChatGPT, Cursor or your own agent.

  2. Your MCP server

    OAuth 2.1 sign-in, scoped tools, limits and a log of every call.

    • search
    • get
    • create
    • update
    • approve
  3. Your application and data

    The same business rules and permissions as your own interface.

The agent only sees the tools you expose, with the rights of the person who connected it.
What we can build

What we can build

From a read-only connector for your knowledge base to an agent that runs whole workflows in your product.

MCP servers for your product

A remote MCP endpoint (JSON-RPC over streamable HTTP) that exposes your app's features as well-described tools.

OAuth 2.1 authorisation

Discovery, dynamic client registration, PKCE, consent screens, short-lived access tokens and rotating refresh tokens stored only as hashes.

Permissions and guardrails

Every tool runs with the connecting user's role and plan limits. Risky actions need explicit human approval, and the server refuses whatever breaks your rules.

Standing instructions

On connect, the server hands the agent your playbook: how to work, which rules apply and what it must never do.

Audit logs and monitoring

Every call is logged with who, what and when, so you can trace exactly what an agent did and why.

Agent workflows

Multi-step work across tools, like researching, drafting, reviewing and queueing, with a person signing off at the right moments.

Proof, not promises

Two MCP servers we run in production

Our sales system runs on an MCP server with 35 tools. Claude handles prospect research, data enrichment, email verification, drafting and reply handling, and a person approves every email before it's sent.

Salestreak has an MCP connector for online sellers with 37 tools for catalogue, stock, photos, print on demand, webshops and orders. The agent works with the role of the user who approved it and within the limits of their plan.

  • Our own OAuth 2.1 servers, with dynamic client registration and audience-bound tokens (RFC 8707)
  • Access tokens that expire after 24 hours and refresh tokens that rotate on every use
  • Approval tools that only run when a person explicitly asks, with the reviewer recorded
  • Server-side checks that refuse drafts or actions that break the rules, whatever the agent says
  • Hundreds of automated tests covering the MCP and sales tools

A working day for our sales agent

  1. Find and research

    Search the CRM, look up companies and save research notes.

    • search_companies
    • save_research
  2. Verify

    Check that an email address exists before any effort goes in.

    • verify_email
  3. Draft

    Write a personal email that follows the campaign brief.

    • create_outreach_draft
  4. Human approval

    Nothing is sent until a person has approved it.

    • approve_outreach_drafts
  5. Handle replies

    Classify replies and draft answers. An opt-out is suppressed everywhere at once.

    • classify_reply
    • draft_reply
Tool names as the agent sees them on our sales MCP server.
MCP servers in production
2
tools across both servers
72
with PKCE and dynamic registration
OAuth 2.1
of outgoing emails approved by a person
100%
Who it's for

Who needs an MCP server

SaaS companies

Let your customers work with your product from Claude or ChatGPT, right next to their other tools.

Teams with internal systems

Give your team an assistant that searches the CRM, updates records and prepares work in your own systems, with the permissions they already have.

Operations and support

Agents that look up orders, customers and stock, and prepare actions for a person to confirm.

Sales and growth teams

Research, enrichment, drafting and follow-up handled by an agent, with every outgoing message reviewed by a person.

How it works

How we build an MCP server

An MCP server is an API designed for an AI. The design work is in deciding what that AI may do.

  1. 1

    Pick the jobs, not the endpoints

    Which tasks should an agent handle end to end? We design the tools around those tasks, with names and descriptions an AI understands.

  2. 2

    Decide what needs a human

    Reading is usually fine. Sending, spending and deleting get approval steps and hard limits on the server.

  3. 3

    Build the auth properly

    OAuth 2.1 with PKCE, dynamic client registration, consent screens and tokens bound to the right workspace and user.

  4. 4

    Write the playbook

    Standing instructions the server sends on connect, so every session starts with your rules and your tone of voice.

  5. 5

    Test with real agents

    We run the tools through Claude and other clients, and add server-side checks wherever an agent could get it wrong.

  6. 6

    Log, measure, extend

    Audit logs, usage and error monitoring, then new tools as your team finds more work to hand over.

Protocols and tooling

The open Model Context Protocol, OAuth 2.1 and the same application stack as the product itself, so the agent follows exactly the same business rules.

  • Model Context Protocol
  • JSON-RPC
  • Streamable HTTP
  • OAuth 2.1
  • PKCE
  • RFC 8707
  • Claude
  • ChatGPT
  • Ruby on Rails
  • PostgreSQL
Cost & timeline

What does an MCP server cost?

Less than most people expect when your application already has clean business logic, more when it doesn't. The estimate depends on:

Number and depth of tools
A handful of read-only tools is a small project. Dozens of tools that change data need more design and testing.
Authentication
Reusing existing accounts through OAuth 2.1 takes more care than an internal API key, but it's what lets customers connect safely.
Approval flows
Human sign-off, limits and refusals for risky actions add work, and they're what makes an agent safe to use.
State of your application
Clean, well-tested business logic can be exposed quickly. Logic buried in old code or spreadsheets needs untangling first.
Clients to support
Claude, ChatGPT, Cursor or your own agent: every client has its own quirks to test against.
FAQ

Questions about MCP servers and AI agents

What is an MCP server?

MCP (Model Context Protocol) is an open standard that lets AI assistants such as Claude and ChatGPT use external tools. An MCP server exposes features of your software as tools the AI can call, like searching records or creating a draft, within the permissions you set.

How much does it cost to build an MCP server?

It depends on the number of tools, whether they only read or also change data, the authentication and the approval flows you need. After a free intro call you get a written scope and estimate.

Is it safe to let an AI agent into our systems?

It can be, if the server is built for it. Our servers use OAuth 2.1, give the agent the same rights as the person who connected it, require human approval for risky actions and log every call.

Which AI assistants can connect?

Any client that supports remote MCP servers, including Claude and ChatGPT, developer tools such as Cursor, and agents you build yourself.

Do we need an API first?

No. We can build the MCP server directly on your application's business logic. If you already have an API, the MCP tools can build on top of it.

Can you build an AI agent that does real work, not just chat?

Yes. Our own sales agent researches, verifies, drafts and handles replies through MCP tools every day, with a person approving every email it sends.

Want AI agents working in your software?

Tell us what you'd like an agent to do and which systems it should touch. You'll get an honest view on what's safe to automate and what should stay with people.

martin@balk.dev

Typically a reply within one business day.

Useful to mention

  • What you want to build, in a few sentences
  • Who is going to use it
  • Any deadline or launch date
  • Tools and systems it should connect to