Teams rolling out AI agents all hit the same wall: the agent is capable, but it cannot reach the tools the business runs on. The ERP, the internal API, the legacy system holding the data that matters. This post explains how MCP solves that, how to decide whether to build, buy, or auto-generate an MCP server, and how superglue lets you build, run, and maintain custom MCP servers without a developer team.
What is MCP (Model Context Protocol)?
MCP is an open standard, introduced by Anthropic in November 2024, that connects AI agents to tools and data. Instead of writing manual connections for every agent-to-tool pairing, you expose a tool through an MCP server, and any MCP-aware agent can use it in a consistent way. Think of it as a common plug: the agent speaks MCP, your tool speaks MCP, and they interoperate without custom wiring on both ends. The full specification is open at modelcontextprotocol.io.
What is a custom MCP server?
Public MCP servers already exist for common products, and vendors ship official ones for tools like GitHub, Slack, and Stripe. But the standard MCP server often exposes too little, too much, or the wrong data for your use case.
A custom MCP server is one you build yourself: for an internal API, a legacy ERP, a proprietary database, or a workflow specific to your business. It exposes those capabilities to AI agents in the MCP format, so the agent can query the system, trigger actions, or read data through a standard interface rather than a one-off integration.
For most enterprises, the tools worth exposing to an agent are the ones no vendor has built an MCP server for, because they are internal. That is why custom is the common case.
When should you build a custom MCP server?
Build one when three things are true:
- You have an AI agent that needs your data or actions to be useful.
- The systems it needs are internal or legacy, with no existing MCP server.
- The value of the agent depends on reaching them reliably.
A support agent that cannot read your ticketing system, or an analyst agent that cannot query your ERP, is a demo. The custom MCP server is what turns your agent into something people use.
Build, buy, or auto-generate?
There are three paths, and the right one depends on your team.
| Approach | Effort | Maintenance | Control |
|---|---|---|---|
| Use a standard MCP server | Minutes to connect | Vendor-owned, on their schedule | Fixed scope; you get what it exposes |
| Build by hand | Roughly two developer-weeks for a first production MCP server (auth, error handling, schema mapping, deployment) | Yours, forever; every API change is a ticket | Full, if you know how to code |
| Auto-generate with superglue | Minutes from a natural-language explanation | superglue detects source changes and proposes fixes | Full; you review every tool before it goes live |
If a maintained public MCP server fits your tool, use it. If the system is internal, the choice is between hand-building and generating. Generating cuts build time by 98% compared to hand-writing the MCP server, with every tool reviewed before it goes live.
How do you build an MCP server in superglue?
You tell superglue what you need in plain English, like you would explain it to a colleague. For example: "Pull my customer list from Salesforce into BigQuery." No code required. The superglue agent works out how to connect the systems, tests the connection, and confirms each step with you before saving. Once you have working connections, you bundle them into an MCP server: pick which ones to include, give the MCP server a name, and superglue hosts it. From that point on, any AI assistant that supports MCP, including ChatGPT, Claude, Gemini, and Copilot, can use those connections directly, with no setup on their end.
Then, superglue keeps the MCP server running and proposes a fix when the underlying system changes, so a renamed field or a versioned API does not silently break the agent. You describe the tool; superglue builds, runs, and maintains the custom MCP server around it, no platform team required.
How do you deploy the MCP server to Claude, Langdock, or ChatGPT?
Once superglue hosts your MCP server, you download the ready-made configuration file and drop it into the AI your team uses:
- Claude: add the MCP server as a custom connector in Claude's settings, paste the MCP server URL, and your agents can use the tools right away.
- Langdock: add the MCP server in your Langdock workspace integrations, and every assistant in the workspace can reach your internal systems.
- ChatGPT, Copilot, Cursor: any MCP-aware client connects with the same config; one MCP server serves all of them.
Deploying to a new client means pasting a URL, not writing code.
Who maintains an MCP server once it's live?
An MCP server is an integration, so it inherits every integration's upkeep problem: the underlying API versions, auth expires, the schema drifts, and the MCP server has to keep pace or the agent silently loses a capability. If you hand-build ten MCP servers, you have ten things to maintain, and the maintenance does not stop.
This is where superglue takes over. superglue builds the custom MCP server, runs it, and proposes a fix when the source system changes, so the MCP servers do not quietly rot the first time an internal API is updated. Customers cut integration maintenance overhead by 80%.
A decision checklist
- Is there an AI agent whose value depends on reaching this system? If no, wait.
- Does a maintained public MCP server already exist for it? If yes, use it.
- Is the system internal or legacy? Then a custom MCP server is the path.
- How many such systems do you have? Hand-building costs developer-weeks per MCP server; superglue pays off from the first one.
- Who owns maintenance when the underlying tool changes? Someone has to; with superglue, that someone can be you, no platform team needed.
A custom MCP server connects your AI agents to the systems your business runs on. With superglue, it is easy to build and maintain: describe what you need in plain English, and superglue generates the MCP server in minutes, hosts it, and proposes a fix when the source system changes.
FAQ
Is MCP only for Claude?
No. MCP is an open standard. Claude, OpenAI's Agents SDK, Microsoft Copilot Studio, Cursor, and most agent frameworks support it, so one MCP server works across agents.
What is the difference between an MCP server and an API?
An API exposes your system to developers; an MCP server exposes it to AI agents. The MCP server wraps your API (or database, or legacy system) in a format agents can discover and call without custom integration code.
How long does it take to build a custom MCP server?
Hand-built: roughly two developer-weeks for a first production MCP server, plus ongoing maintenance. Generated with superglue: minutes from a natural-language explanation, with hosting and maintenance included.
How do I connect a custom MCP server to Langdock or Claude?
Download the MCP server config from superglue and add it in the client's settings: as a custom connector in Claude, or under workspace integrations in Langdock. The same MCP server works in both.
Do I need developers to build one?
Not with superglue. You describe what data should move where in plain English; superglue builds and tests the connections, confirms each step with you, and hosts the bundled MCP server.
Build your first custom MCP server
Describe what you need in plain English and your MCP server is live in minutes.
Get started