To monitor an MCP server in production, test it from outside the way an agent calls it. On a schedule, connect to the server, read its tools list, call one tool and check the result. Tracing inside the server shows why a call failed. Only an outside test shows that agents cannot connect at all.
This guide is for teams that ship an MCP server to customers, next to their API. The examples use Fixter (fixter.dev), which is our product, and every claim about it is cited to its docs, checked on November 13, 2026. If you are looking for MCP servers that let Claude Code read your own logs and errors instead, that is a different question: see monitoring MCP servers for Claude Code.
Tracing inside the server, testing from outside
Most guides to MCP server monitoring cover one view: tracing inside the server. You add an SDK, and every tool call that reaches your code becomes a span with its name, duration and error. Sentry's MCP monitoring works this way, through its JavaScript and Python SDKs.13 It is the right tool for the question "why did this call fail".
It has one blind spot. A trace starts when a request reaches your code. If the request never arrives, or the agent never sends it because your tool vanished from the list, nothing is recorded and nothing alerts. For an MCP server that is the common way to break: the process is healthy and your customer's agent still cannot use it.
| What broke | Tracing inside the server | A scheduled test from outside |
|---|---|---|
| DNS, an expired certificate, or a deploy that left the server unreachable | Nothing. No request arrives, so there is no span and no error. | Fails on the next run, with the cause: DNS lookup, TLS handshake, could not connect, or timed out. |
| A gateway, CDN or auth layer in front rejects clients | Nothing. The request is stopped before your code. | Fails: the server did not identify itself, or refused the credential. |
| A release drops or renames a tool | No error. Agents stop calling the tool, so traffic just goes quiet. | Fails and names the tool that is missing from the list. |
| The server stops accepting a protocol revision that older clients speak | Usually nothing you would alert on. | A test pinned to that revision fails and names it. |
| Slow for the client: DNS, TLS, the network, then your handler | Sees the time spent in your handler only. | Measures the whole session and fails over the limit you set. |
| Broken at 3 am, when no customer is calling | Silent until the first real call fails. | Found on the next scheduled run. |
| A tool throws on one customer's input | Yes, with the stack trace and the arguments. | Only if the test happens to call that tool in a way that triggers it. |
| Which line, query or downstream call was at fault | Yes. This is what tracing is for. | No. It tells you the call failed, not why. |
The outside column describes Fixter's MCP tests.1,3 The two views do not replace each other: the last two rows are why you still want tracing.
What an agent does, and what to check at each step
An agent uses an MCP server in three moves. It connects and finds out what the server speaks: on revisions up to 2025-11-25 that is an initialize handshake, and on the current revision, 2026-07-28, there is no handshake and the client may ask with server/discover first.12 It asks which tools exist, with a tools/list request. Then it calls one, with tools/call.11 A good test walks the same path and lets you choose how far a healthy server has to get.
Fixter finds and fixes the issues in your app in production, from your coding agent: errors, failing infrastructure, slow requests and downtime. It finds the root cause and fixes it: in your agent over MCP, or as a pull request ready for your review. Nothing to learn, built for small teams. Its uptime monitoring calls your HTTP endpoints and your MCP servers from outside on a schedule.2 An MCP test has four check depths, and each one includes the ones before it.1
| Check depth | What the test sends | Passes when | Use it to catch |
|---|---|---|---|
| Uptime | The handshake | Something at the URL answers as an MCP server. A 401, 402 or 403 counts as answering. | The server is gone or unreachable. Needs no credential. |
| Handshake | The handshake | The server identifies itself and the handshake completes. | A server that is up but turns clients away. |
| Tools list | The handshake, then tools/list | Every tool you expect is in the list. | A release that dropped or renamed a tool. |
| Tool call | The handshake, tools/list, then one tools/call | The tool you name returns a result you accept. | A tool that is listed but broken, or answers with the wrong shape. |
From Fixter's MCP tests docs.1 Every depth can also fail on a latency limit.
The tool call is the one that matters most, because of how MCP reports errors. A tool that fails does not have to return an HTTP error. The spec has two kinds of failure: protocol errors, sent as JSON-RPC errors, and tool execution errors, sent as a normal result with isError: true.11 A check that only looks at the status code sees a healthy 200 in the second case. Fixter's tool call test fails on any result marked isError: true, whatever else you check.1
You can also check the shape of the result. Give the test a sample such as {"results": []} and the run fails when the tool stops returning a results array.1 That catches the quiet failure where the tool answers, with something an agent can no longer use.
Set it up from your coding agent
You create a test by asking your agent. The commands below are for Claude Code; any MCP client that can connect to Fixter works the same way.
- Connect Fixter. In a terminal, add Fixter's MCP server and sign in.9
- Start with a test that needs no credential. An Uptime test only asks whether the server answers, so it works against a server that wants a token before it says anything.1
- Store a credential for the deeper checks. Tools cannot be listed without getting in. Fixter takes a bearer token, an API key, basic auth or OAuth2 client credentials.6 Ask your agent to store it, or create it in the app if you would rather not paste a token into a chat.7 Use a token made for the test, with read access only. The spec lets a server show only the tools a token's scopes allow, so give the test token the scopes your customers' tokens have, or list only the tools it can see.11
- Ask for the test you want, in plain language. Name the tools that must be listed, the tool to call and its arguments.
{{uuid}}sends a fresh value on every run, so the call never depends on data an earlier run left behind.1 - Let the save prove it. Saving a test runs it once first, and a test that fails that run is not saved.3 So a typo in a tool name shows up now, not as a false alarm at night.
- Choose where failures go. By default they go to your email. Connect a Slack channel if you want them there too, and ask your agent to route this test to it.4
Pick a read-only tool for the call: a search, a list, a status. The test runs on every interval, so a tool that creates an order would create one every minute. If a CDN or bot protection sits in front of your server, allow the test's User-Agent, or the failures you see will be your firewall and not your server.8
What you see when it fails
A test runs every 60 seconds unless you change it, and every 30 seconds at the shortest. One failed run does not page you: the test goes down after three failed runs in a row by default, and you can set that from 1 to 10.3 You get one message when it goes down and one when it recovers, by email or Slack, not one per run.4
The message says what failed in words, not in codes.1,4
MCP server did not identify itself: nothing at the URL answers as an MCP server.MCP server refused the credential: the token the test uses stopped working, which usually means your customers' tokens are about to.Tool get_order is missing: the server is up, and a tool agents depend on is gone from the list.Tool search_orders returned an error: the tool is listed and broken.Took 2400 ms, limit 2000 ms: it works, too slowly.
Each run also records uptime and latency, and where the time went: DNS, connect, TLS, the server, the download.3 Ask your agent "what was the uptime of the orders MCP server this week" and it reads the numbers back.
This is where the two views meet. Every run sends its own trace id to your server in a traceparent header. If your MCP server sends its traces to Fixter, the failed run links to what happened inside it: the slow query, the error from a downstream API.3 You get told that agents cannot use the tool, and why, in one place. Fixter stores logs, traces and metrics sent over standard OpenTelemetry, so the inside view needs no Fixter SDK. For an API next to your MCP server, the same tests work on plain HTTP endpoints: see API monitoring for a small team.
Keep older clients working
MCP is versioned by date. The current revision is 2026-07-28, and a server may support several revisions at once.12 Your customers do not all upgrade their clients on the day you upgrade your server, so "which revisions do we still accept" is a production question, and one an SDK upgrade can change without anyone deciding it.
By default a Fixter test speaks whatever revision the server answers with and never fails over it. To hold a revision in place, pin a test to it: create one Handshake test per revision you promise to support. When a release drops one, only that test fails, with Server no longer speaks protocol 2025-06-18.5
Where an outside test falls short
- It does not tell you why. A failed tool call says the tool returned an error. The cause is in your server's logs and traces. Keep tracing; link the two with the trace id.
- It covers the path you test. One test calls one tool with one set of arguments. An error that only one customer's data triggers will not show up in it.
- Public URLs only. The test needs an http or https URL reachable from the public internet. Private addresses are refused, and a server that only runs locally over stdio cannot be tested.3
- One location. Fixter's tests leave from one static address, in region eu-west-1.8 You learn that the server is slow or down from there, not from every continent.
- No browser sign-in. The test authenticates with a stored secret. A server that only accepts an interactive sign-in needs a token you issue for the test, or an Uptime test that checks the server answers.6
On cost: one run is one check. A test every minute is about 43,000 checks a month, and the free plan includes 50,000 a month. Paid plans start at €20 a month.10 If you are deciding how to hear about downtime in general, start with how to know when your app goes down.
Key takeaways
- Tracing inside an MCP server only sees requests that arrive; a server nobody can reach produces no errors
- Test from outside in the order an agent works: connect, tools/list, one tools/call
- A broken tool can answer 200 with isError: true, so check the result, not the status code
- Check that the tools you promise are still listed after every release
- Pin one test per protocol revision you still support
- Keep tracing for the why, and link it to the test with the trace id
Frequently asked questions
How do I monitor an MCP server in production?
Use two views. Trace the server from inside to see why a tool call failed. Test it from outside on a schedule to see whether an agent can use it at all: connect the way a client does, read the tools list, call one tool and check the result. The outside test is the one that catches a server nobody can reach.
What is the difference between tracing an MCP server and testing it?
Tracing runs inside the server and records the requests that arrive: which tool, how long, what error. Testing runs outside and sends its own requests on a schedule. Tracing cannot see a request that never arrived, such as one stopped by DNS, a certificate or a gateway. A test cannot see the line of code that failed. They answer different questions.
Can I test an MCP server that requires authentication?
Yes. In Fixter an Uptime test needs no credential: a 401, 402 or 403 counts as the server answering. For the handshake, the tools list or a tool call, store a credential first: a bearer token, an API key, basic auth or OAuth2 client credentials. There is no browser sign-in, so the test needs a secret it can send on its own.
How often should I test my MCP server?
Every minute is a sensible default for a server customers depend on. Fixter runs a test every 30 seconds at the shortest and marks it down after three failed runs in a row unless you change that. Call a read-only tool, because the test runs on every interval.
Can I test a local MCP server that runs over stdio?
No. The test needs an http or https URL that is reachable from the public internet. A server that only runs on your machine over stdio, or on a private network, cannot be reached, and private addresses are refused.
What does it cost to monitor an MCP server with Fixter?
One run is one check. A test that runs every minute uses about 43,000 checks a month, and Fixter's free plan includes 50,000 checks a month, so one test every minute fits on the free plan. Paid plans start at €20 a month.
Sources (13), all checked November 13, 2026
- Fixter docs, MCP tests (check depths, fields, what a failed run says)
- Fixter docs, API & MCP testing overview
- Fixter docs, runs, states and uptime (interval, failure threshold, uptime, trace id, public addresses only)
- Fixter docs, API test notifications (email, Slack, what each message says)
- Fixter docs, MCP protocol revisions (policies, pinned tests)
- Fixter docs, credentials (the four types)
- Fixter docs, creating your first test (with your agent, in the app)
- Fixter docs, allow tests through your firewall (User-Agent, IP address, region)
- Fixter docs, MCP setup (Claude Code command)
- Fixter pricing (checks per plan, what one check is)
- Model Context Protocol specification, tools (tools/list may vary by authorization, tools/call, error handling)
- Model Context Protocol specification, versioning (current revision, server/discover, supporting several versions)
- Sentry docs, set up MCP server monitoring (SDK instrumentation inside the server)