MCP and WebMCP
- MCP
- Model Context Protocol. A server exposes your product’s actions as tools to clients outside the browser, such as ChatGPT, Claude or a coding agent, with its own authentication.
- WebMCP
- A proposed web standard. Your web page registers tools, and an agent working in that page uses them with the session the user is already signed in with.
An API is not automatically a good agent interface. Endpoints are often too broad, overlap in meaning, or expose actions that need additional confirmation. A model picks tools by their names and descriptions, fills in arguments from context and retries when a call fails, so wrapping every endpoint as a tool rarely works. The work is deciding which actions exist, what each one may do, and what the system checks regardless of what the model asks for.
Typical scope
- Tool inventory: which actions an agent needs, and which it should never have
- Tool schemas: names, arguments and descriptions
- Authentication and scopes, tied to the user the agent acts for
- Confirmation before destructive actions
- Idempotency
- Error design: what went wrong, and what to try next
- MCP and WebMCP implementation
- Tool evals and regression tests
Engagements
Ongoing
Not included
- Changes to your core API beyond what the interface needs (quoted separately)
- Building the agent itself (see AI agent architecture)
- Hosting: the server runs in your infrastructure
Questions
How much does an MCP server cost to build and run?
An agent interface audit starts at €1,500. An MCP server or WebMCP interface for a defined set of actions starts at €4,000, and a production interface with confirmation flows, audit logging and evals in your CI at €8,000. Running it costs what your infrastructure charges, because the server runs there under your accounts. The agent reliability retainer, from €750/month, keeps it working as models and your API change.
Why build an MCP server for our product?
So the assistants your customers already use can act in your product with the permissions of the user they act for: create a record, update a ticket, check an order. Without one, an assistant has to work from your UI or your public documentation, or it cannot do the task at all.
Do we need MCP or WebMCP?
If customers use assistants away from your site, MCP. If agents act inside your web app, WebMCP. Many products need both, sharing the same tool contracts.
How does authentication work for an MCP server?
Remote MCP servers use OAuth: the client gets a token for a specific user, and the server checks it on every call. I tie scopes to the user the agent acts for, and to the customer workspace where customers share one server, so a tool can never do more than that user could do in your product. WebMCP tools run in your page and use the session the user is already signed in with.
How do you test MCP and WebMCP tools?
Evals are tasks a model should complete with your tools, checked for the right calls and arguments. Regression tests cover scopes, confirmations and errors. Both run in your CI, so a changed description, schema or model version that breaks a tool fails there first.
Can our own frontend team add WebMCP?
Yes. Registering a tool in a web page is ordinary frontend work. Where I help is deciding which actions to expose, describing them so a model picks the right one, and placing confirmations before anything is bought, sent or deleted.
Writing
MCP & WebMCP
Discuss a project
What you are building, what is going wrong, and roughly when. I answer within a couple of days.