We have built twelve MCP connectors on our own products, from a seven-tool roster connector to a thirty-two-tool league manager. This is what they cost, where the time goes, and what we would tell anyone pricing one.
What an MCP server is, in one paragraph
MCP stands for Model Context Protocol, the open standard ChatGPT and Claude use to connect to outside software. An MCP server describes a set of tools ("show this week's bookings", "publish the schedule", "change this price"), and the assistant calls them when a user asks for something. Your server checks who the user is, does the work against your real data, and sends back the result. To the user it feels like ChatGPT suddenly knows their account.
Build cost by scope
| Scope | Typical tools | Build cost (CAD) | Time |
|---|---|---|---|
| Read-only, public data | 5 to 10 lookups, no sign-in | $3,000 to $6,000 | 1 to 2 weeks |
| Focused, signed-in | Read tools plus a few confirmed writes, OAuth | $4,000 to $12,000 | 2 to 3 weeks |
| Full product connector | 15 to 30+ tools, several roles, previews before every write | $12,000 to $30,000 | 3 to 6 weeks |
US clients can read those as roughly 0.72 of the CAD figure. Directory review is on top of the build time and outside anyone's control.
Where the time actually goes
- Sign-in: OAuth with dynamic client registration and a consent page on your domain, so each user connects their own account. This is usually the biggest single piece the first time.
- Permissions: every tool has to check that the user can see that league, store or record. Getting it wrong leaks one customer's data to another, so it gets its own tests.
- Write tools: anything that changes data previews the change first and waits for the user to confirm. Overwrites and deletes are marked as destructive so the assistant handles them carefully.
- Review package: test accounts, example prompts that should and should not trigger tools, listing copy and icons for OpenAI's and Anthropic's directories.
- Tests: a suite that runs every tool against a demo account, so later changes to your app do not quietly break the connector.
What it costs to run
Less than most people expect. The model runs on the user's own ChatGPT or Claude plan, so you are not paying per question. Your cost is hosting the server, which usually lives inside the app you already run, plus the database reads each tool call makes. For our products that has been close to zero on top of existing hosting.
The ongoing cost is maintenance: when your app changes, the tools that touch that part need updating, and assistants occasionally change how they call tools. Plan on a few hours a quarter for a stable product.
Is it worth it?
It is worth it when your users already live in ChatGPT or Claude and your product has a few jobs they repeat: checking status, entering results, booking, repricing. A connector turns those jobs into one sentence. It is not worth it as a novelty on a product people use once a year.