An MCP connection can create several bills. The server operator, the underlying service, and the AI provider may each charge for a different part of the work.
That makes “free MCP server” an incomplete budget description. It may mean the source code is available without a license fee, that hosting has a free allowance, or that an existing subscription includes the connector. Establish which meaning applies before relying on it.
Follow a task through its dependencies
Suppose an assistant searches an order system and summarizes an order's status. The request needs an AI client, an MCP server, and access to the order system. If the workflow sends a notification, it also uses a delivery service and needs permission to act.
List those dependencies before comparing vendors. For each one, record the payer, billing unit, included allowance, and behavior when a limit is reached. Include the person responsible for maintenance. Running an open-source server yourself makes you the operator.
A simple worksheet can track:
- Server hosting and storage.
- Model or client subscription costs.
- Paid APIs and data subscriptions.
- Support, monitoring, updates, and incident handling.
- Data transfer or other usage charges where applicable.
Mark only the categories that apply to the server and services you choose.
Read what “free” includes
A source repository can provide software you are allowed to run while leaving you responsible for its hosting, dependencies, and maintenance. A hosted free tier can remove that installation work but impose its own usage or feature limits. A connector bundled with a subscription may still require a separately paid account in the system it connects to.
Use the order lookup to make this concrete. Name the assistant subscription, server operator, and order-system account. Ask whether the server adds a fee, passes through another provider's charges, or relies on services you must supply. If a model is called inside the server as well as by the assistant, include both paths.
Then examine failure behavior. Does a limit refuse new requests, reduce throughput, or create another charge? Can an operator stop usage? How does a customer see which requests were billed? Get those answers from the actual terms and implementation. “Free” by itself cannot tell you how an unattended workflow behaves when it reaches a limit.
For a personal experiment, setup time may dominate the decision. For a business workflow, support ownership and recovery from a failed task may matter more. Use the cost categories that fit the way you intend to run the service.
Compare free and paid against the same task
Use a representative task and record whether it completes correctly. Check what happens when the upstream service is unavailable, an authorization expires, or the response is too large.
A free allowance may be enough for learning or occasional personal use. A team may value a paid service's documented support or operational ownership. Read the terms rather than inferring those features from the word “paid.”
For usage-based services, count completed tasks and retries during an evaluation. A cheap request that repeatedly fails can create more work and more billable activity than the initial price suggests. Keep the model bill separate from the server bill so you can identify which part changed.
OBTO's pricing page describes its application-based plans and how separately paid AI usage fits. Verify the current terms for your workload; this article does not quote a fixed price that may become stale.
An evaluation log should make it possible to reconcile the server bill with the model bill. If a retry appears in both, count both charges when estimating the cost of a completed task.