Description
Check whether what people use the tool for matches the jobs it is meant to support.
MCP tool analytics shows which tools people use, which fail the most, which users hit errors, how slow each tool is and its call success rate. Flowlines adds what people were trying to do with each tool and notifies your team in Slack, by email or through a webhook when failures recur.
100 analyzed MCP sessions per month, free. No credit card. Connect your assistant, then instrument your server.
Select a use case to see what people were trying to do with the tool and where it failed.
| Tool | Calls | Users | Success | p95 |
|---|---|---|---|---|
search_customers | 8,214 | 342 | 99.3% | 410 ms |
list_invoices | 5,108 | 211 | 97.1% | 880 ms |
create_refund | 1,302 | 96 | 99.8% | 702 ms |
update_plan | 884 | 120 | 94.2% | 640 ms |
export_pdfNever called | 0 | 0 | n/a | n/a |
Jobs behind search_customers
MCP tool analytics measures how each tool exposed by an MCP server is used and how reliably it works: calls, users, failures, latency and the jobs behind those calls. Flowlines shows this per tool, next to what people were trying to do with it.
The same operation can serve very different jobs. search_customers may support renewal preparation, account lookup or duplicate-record investigation, and each can fail for a different reason.
Flowlines ranks your MCP tools by the users and sessions that call them, and shows what people use each tool for.
A tool called 300 times by one user is a different story from a tool called 30 times by 25 users. Start with reach, then look at volume and use cases.
| Measure | Use it to | Avoid concluding |
|---|---|---|
| Users per tool | See which tools reach the most people. | Every user finished the job the tool supports. |
| Sessions per tool | Separate broad adoption from repeated use by a few people. | Every call represents a different customer. |
| Calls per tool | See where activity concentrates. | More calls always mean more value. |
| Use cases per tool | Understand the work the tool supports. | A tool name alone reveals the user's goal. |
Flowlines ranks your MCP tools by failed calls, with the users and sessions each failure affects, so the tool that hurts the most people rises to the top.
Compare failure rates, failed calls divided by calls with a known result, and require a minimum number of calls; raw failure counts favor the busiest tool. Split failures by class too: a downstream timeout and a validation error from a vague argument need different owners.
| Failure class | Example | Usual owner |
|---|---|---|
| Transport or authentication | Expired token or rejected connection. | Platform or auth. |
| Protocol validation | Arguments that do not match the schema. | Tool design: description and schema. |
| Tool error | A result returned with isError: true. | The tool's domain logic. |
| Downstream dependency | The API behind the tool timed out. | The dependency owner. |
| Timeout or cancellation | The client gave up waiting. | Latency work; report separately from errors. |
Flowlines lists the users behind a tool's failed calls, so you can see whether one customer keeps hitting the same error or many people hit it once.
The two cases call for different responses. One user failing repeatedly is a support conversation and possibly an account-specific bug. Many users failing once each points to the tool itself. Open an affected session to see what the person was trying to do when it failed.
| Pattern | Likely meaning | Next step |
|---|---|---|
| One user, many failures | Account data, permissions or a usage pattern. | Contact the user and inspect their sessions. |
| Many users, one failure each | A defect in the tool or a dependency. | Fix the tool; check the failure class. |
| One client only | A client-specific integration problem. | Compare the same job in another client. |
MCP tool call success rate is successful calls divided by calls with a known result in the same window, and Flowlines shows the calls and failures behind it for every tool.
In the synthetic example on this page, search_customers has 360 calls and 12 failures: 348 / 360, or 96.7%. The rate differs by use case, and the smallest one has the lowest success rate. Report unknown results separately, and never call this task completion: a call can succeed while the user's job stays unfinished.
| search_customers use case (synthetic) | Calls | Failed | Call success rate |
|---|---|---|---|
| Prepare a renewal brief | 180 | 4 | 97.8% |
| Find the right account | 110 | 3 | 97.3% |
| Investigate duplicate records | 70 | 5 | 92.9% |
| All use cases | 360 | 12 | 96.7% |
Flowlines shows how long each tool takes next to its failures and use cases; measure the p95 duration per tool rather than the average, because averages hide the calls people notice.
Compare the same tool on the same server and environment, and break a slow tool into the downstream calls it makes. Report timeouts and cancellations separately from errors. A slow tool matters most when it stops people from finishing their work, so read latency with the sessions it affects.
| Measure | What it tells you | Watch out for |
|---|---|---|
| p50 duration | The typical call. | A stable median can hide a slow tail. |
| p95 duration | The calls users notice. | Small samples make high percentiles unstable. |
| Latency by dependency | Which API or database is slow. | Duplicate spans inflate counts and timing. |
Flowlines notifies your team in Slack, by email or through a webhook when tool failures recur, after analysis, with the affected users and sessions attached.
For paging on a numeric threshold within minutes, such as more than 5% errors over 5 minutes with at least 100 calls, keep a rule in your metrics backend. The two work together: the threshold says something broke, and Flowlines shows who it affected and what they were trying to do.
| Alert | Fires on | Best tool |
|---|---|---|
| Recurring tool failure | A failure pattern across users and sessions. | Flowlines, to Slack, email or a webhook. |
| Error-rate threshold | A percentage over a short rolling window. | Your metrics backend. |
| Slow tool | p95 duration above your objective. | Your metrics backend. |
Flowlines shows observed activity for every tool it has seen, so a tool with no calls in the selected period stands out; confirm your tool catalog and telemetry coverage before calling it unused.
Zero calls does not tell you why. The tool may be seasonal, inaccessible to a client, poorly described or simply unnecessary. A declared tool is not proof that a particular client discovered it. Validate those hypotheses before removing or rewriting it.
| What you observe | Possible explanation | Check next |
|---|---|---|
| Declared, zero calls | No demand, unclear description or not discovered. | Read the description as a model would; test it in each client. |
| Zero calls for every tool | Telemetry gap, not a usage drop. | Exporter health and a known test call. |
| Calls from one client only | Discovery or permission differs by client. | Tools by client and client-specific errors. |
Flowlines compares tool usage across time windows and shows contract changes, so you can check a description or schema change against the sessions it affected.
Use equal windows with comparable traffic. A lower error rate with fewer users or less outcome coverage is not automatically an improvement. Treat the comparison as an investigation, not a controlled experiment.
Check whether what people use the tool for matches the jobs it is meant to support.
Look for changed arguments, failed calls and repeated attempts around the new version.
Compare the path people take and how their sessions ended, not just the number of calls.
The Flowlines plugin gives your assistant the setup workflow. Connecting the assistant does not automatically instrument your MCP server.
MCP OpenTelemetry instrumentation guide ↗MCP observability product overview ↗The MCP analytics measurement guide ↗
Flowlines ranks tools by failed calls, with the users and sessions each failure affects. Compare failure rates with a minimum number of calls, and split failures by class before assigning an owner.
Flowlines lists the users behind a tool's failed calls, so you can tell one customer hitting the same error repeatedly from many people hitting it once.
Divide successful calls by calls with a known result in the same window, and report unknown results separately. Label this call success, not task completion. State the time window and exclusions.
Flowlines shows how long each tool takes next to its failures and use cases. Measure the p95 duration per tool and server, not the average, and report timeouts separately.
Flowlines notifies your team in Slack, by email or through a webhook when tool failures recur, after analysis. Keep minute-level threshold paging in your metrics backend.
Flowlines ranks tools by the users and sessions that call them and shows what people use each tool for. A tool called often by one user is not the same as a tool used by many people.
Yes. Use available contract history and comparable analytics windows, then inspect representative sessions. Differences in users, clients or telemetry coverage can affect the comparison.
Start with the traffic you already have
Connect Flowlines and take the next step with your assistant.