Catch a recurring broken path.
Notify the owning team when agents repeat the same operation, recover badly from an error, or declare success without the needed result.
Example: 17 account-enrichment journeys looped after an empty lookup.MCP observability
Understand the usage. Act on what matters.
Connect MCP activity to the people, product needs and business signals behind it. Catch recurring failures and get notified when a session needs your team's attention.
Definition
What is MCP observability?
MCP observability is monitoring how a Model Context Protocol server behaves in production and how people use it: which users and AI clients call it, which tools they use, what they are trying to do, where sessions fail and how fast each tool responds.
Infrastructure monitoring answers whether the server is up and fast. MCP observability adds the usage layer: users, sessions, use cases, outcomes and the signals your product, engineering and sales teams act on. Flowlines provides that usage layer for MCP servers and works alongside Grafana, Datadog and OpenTelemetry rather than replacing them.
Updated . Sources: the MCP specification, the OpenTelemetry MCP semantic conventions and request metadata (SEP-414).
What to monitor
What should you monitor on an MCP server?
Start with eight signals. The first four belong in the monitoring you already run. The last four are the usage layer Flowlines adds: who uses the server, what they want, and what to act on.
| Signal | What it tells you | How to measure it | A starting alert |
|---|---|---|---|
| Availability | Can clients connect and list tools? | Health and readiness probes; failed initialize and tools/list requests | Probe failing for 2 minutes |
| Traffic | How much each tool is used | Calls per second by server, environment and tool | Calls drop 50% against the same hour last week |
| Errors | Which tools fail, including errors returned inside a successful response | Error rate by tool, counting tool results that report an error | Error rate above 5% for 5 minutes with at least 100 calls |
| Latency | Which tools are slow for users | p50, p95 and p99 duration by tool | p95 above 2 seconds for 10 minutes |
| Users and AI clients | Who uses the server, from Claude, ChatGPT, Cursor or others, and who comes back | Users, returning users and sessions by AI client in Flowlines | An important account stops using the server |
| Use cases | What people are trying to do, not just which tool ran | Top use cases and their tool paths in Flowlines | A new use case grows quickly |
| Outcomes | Whether the job got done, beyond a successful call | Sessions that finished, stalled or repeated in Flowlines | A recurring job keeps ending unfinished |
| Product and buying signals | Requests the MCP cannot serve yet, and evaluation questions | Product needs and buying intent in Flowlines | Notify Slack when someone asks about seats or security |
Free starter kit
Copy the alerts and dashboard
Example Prometheus alert rules with tests, a Grafana dashboard for calls, errors and p50/p95/p99 latency by tool, and a production checklist. They use example metric names; adapt them to your own telemetry.
Looking specifically for usage and product analytics? Explore MCP server, tool and session analytics.
01Who uses it
People, agents, accounts, workspaces, and clients when reported with permission.
02What they want
Observed jobs and use cases, not a disconnected list of tool names.
03Where to investigate
Affected sessions, repeated calls, failures, corrections, and reported outcomes that conflict with observed results.
04What to do next
Evidence-backed signals for engineering, product, customer, and sales teams.
Servers and tools
How do I monitor several MCP servers in one place?
Flowlines puts every MCP server in one workspace: which servers are active, which tools drive adoption, which users and AI clients call them, and where reliability changes.
app.flowlines.ai/home
Home
Who uses your MCP servers, where they come from, and what needs a look.
Overview
Users0↑ 18%
Returning users0%↑ 6 pts
Sessions0↑ 32%
Session error rate1.8%↑ 0.4 pts
Live sessionsStreaming
- STSam T.FieldworkChase an unpaid invoice
12s - RLRobin L.OrbitResend a receipt
41s - JMJules M.StudioExplain a failed paymentC1m
- NFNina F.BrightsideFind unpaid invoices2m
Top use casesShare of sessions
Find a customer's unpaid invoices
search_customers → list_invoices27%↑ 4%Issue a refund for a duplicate charge
get_charge → create_refund19%↑ 11%Explain why a payment failed
get_payment → get_events15%↑ 2%Upgrade or downgrade a subscription
get_subscription → update_plan12%↑ 6%Export last month's revenue by plan
list_invoices → export_csv9%↑ 1%FlowlinesAPP#sales-intent · just now
Buying intent · A user at Northstar is evaluating 80 seats
Maya R. asked to roll out to an 80-person finance team with SSO, right after a refund went through in Claude.
Open session 7c1e ›Synthetic data · live demo
Journey reconstruction
A successful tool call is not a successful journey.
Follow the captured calls, repeated attempts and reported result. See the client and permitted identity when supplied. Missing outcome context stays unknown, not automatically successful or abandoned.
MRSession 7c1e · Maya R.Northstar · billing-mcpClaude2.4 s
Claude2.4 sNorthstar was charged twice for order 1932 in March. Find it and refund the duplicate.
- ✓
get_customerNorthstar · cus_8KQ2118 ms - ✓
search_orders3 orders in March342 ms - ✓
get_orderord_1932 · €1,240205 ms - !
list_paymentsSlow · paged twice601 ms - ✓
get_invoiceinv_0457 charged twice231 ms - ✓
create_refund€1,240 refunded702 ms
Thanks. Could we roll this out to our 80-person finance team, with SSO?
Buying intent detected · 80 seats · SSOTop use casesWith tool paths
Find a customer's unpaid invoices
search_customers → list_invoices27%↑ 4%Issue a refund for a duplicate charge
get_charge → create_refund19%↑ 11%Explain why a payment failed
get_payment → get_events15%↑ 2%Upgrade or downgrade a subscription
get_subscription → update_plan12%↑ 6%Export last month's revenue by plan
list_invoices → export_csv9%↑ 1%The missing layer
How does MCP observability fit with Grafana, Datadog and OpenTelemetry?
Keep Grafana, Datadog and OpenTelemetry for server health. Flowlines adds the usage layer above calls and traces: the people, sessions, use cases and outcomes behind them.
| Layer | What it answers | Typical owner |
|---|---|---|
| Infrastructure | Is the service available, fast, and within capacity? | Platform and SRE |
| Tool calls | Which operation ran, failed, or called a dependency? | MCP engineering |
| Agent traces | What did one agent execution do? | AI engineering |
| Usage analytics (Flowlines) | Who uses the server, what are they trying to do, what keeps failing, and who is affected? | Engineering, product, customer, sales |
| Outcomes (Flowlines) | What result was reported, what do the calls support, and what remains unknown? | Product and business |
Flowlines complements Grafana, Datadog, Sentry, Honeycomb, Langfuse, LangSmith, and OpenTelemetry-compatible backends.
Real-time alerts after analysis
How do I get alerted about what MCP users do?
Flowlines sends alerts when it finds buying intent, product needs or broken workflows in MCP sessions. Route them to Slack, email or a webhook as analysis completes, or bundle them in a daily or weekly digest.
See what users ask for next.
Group unsupported requests and manual workarounds by use case, user, and account before deciding what to build.
Example: 31 users asked the MCP to export a report it cannot create.Find friction before a ticket.
Spot an important account repeatedly trying to finish the same job. Use failed resolutions and observed frustration to choose a useful intervention.
Example: A pilot account tried the same setup journey seven times.Review meaningful buying signals.
Surface repeated security, deployment, integration, or scale questions in the right account context. A person decides whether to follow up.
Example: An enterprise account asked about private deployment in four sessions.Sessions with buying intent18+50% vs previous period
RuleNotify #sales-intent when a user asks about pricing, seats or a security review
FlowlinesAPP#sales-intent · after analysis
A user at Northstar is evaluating 80 seats
Maya R. asked to roll out to an 80-person finance team with SSO, right after a refund went through in Claude.
“Could we roll this out to our 80-person finance team, with SSO?”
Claude · session 7c1eOpen session ›
Real-time after analysis, or a daily or weekly digest · Slack · Email · Webhook
Examples are synthetic. Signals are leads for review, not automatic proof of cause or intent.
Delivery follows analysis; product-need grouping can add delay. Explore business KPIs and notification rules ›
From alert to investigation
Open the sessions behind a recurring problem.
See the use case
Understand the real jobs people and agents give the server.
Follow the journey
See typical paths, loops, failures, and reported outcomes.
Find the people
Know which permitted users, accounts, agents, and clients drive the behavior.
Open the evidence
Inspect the relevant request and calls before deciding what changed.
MCP production library
Use the guide that matches your decision.
Choose a production framework, analytics model, tool comparison, build decision, or OpenTelemetry instrumentation guide.
01 / Category guide
MCP observability in production
Build the complete monitoring model across infrastructure, tools, sessions, behavior, and outcomes.
Read the guide ›02 / Product analyticsWhat users do with your MCP
Measure adoption, use cases, journeys, unmet needs, feedback, and commercial signals.
Read the guide ›03 / Buyer's guideThe best MCP observability tools
Compare behavioral analytics, APM, tracing, error tracking, and open-source options fairly.
Read the guide ›04 / Decision frameworkShould you build an MCP server?
Choose between MCP, a direct API integration, an agent skill, static context, and browser use.
Read the guide ›05 / Engineering guideInstrument MCP with OpenTelemetry
Set the data boundary, trace the tool operation, validate failures, and verify production coverage.
Read the guide ›06 / Stack guideAdd MCP analytics to your existing stack
Keep OpenTelemetry, Grafana or Datadog for server health and add users, clients, use cases and signals.
Read the guide ›07 / OpenAI distributionList an MCP plugin in OpenAI
Prepare the endpoint, annotations, tests, policies, review materials, and post-launch measurement.
Read the guide ›08 / Customer storyWhat Melaya learned from its first MCP tool calls
How Melaya uses Flowlines to see which tools people use, from which assistant, and what they try to do.
Read the guide ›09 / Claude distributionList a plugin in Claude
Package, validate, submit, and operate a public Claude plugin with MCP support.
Read the guide ›MCP observability FAQ
The important questions, answered directly.
What is MCP observability?
MCP observability is monitoring how an MCP server behaves in production and how people use it: which users and AI clients call it, which tools they use, what they are trying to do, where sessions fail and how fast tools respond. Flowlines provides the usage side and works alongside the monitoring you already run.
What does Flowlines show for an MCP server?
Flowlines shows observed servers, tools, clients, permitted identities, calls, request patterns, related sessions, outcome coverage and recurring issues. Coverage depends on the integration. An observed session is not necessarily the assistant's complete conversation.
Is MCP latency the main Flowlines metric?
No. Latency and call status remain useful, but Flowlines focuses on what people and agents attempt, where the observed workflow gets stuck and who is affected. Reported outcomes are reviewed alongside the available calls.
Does a missing outcome mean a user abandoned the workflow?
No. Missing outcome context means the result is unknown. A workflow may finish outside the server, or its client may not report a result. Flowlines does not treat silence alone as proof of abandonment.
How is MCP analytics different from MCP monitoring?
Monitoring focuses on availability, latency, errors, and call health. MCP analytics focuses on users, adoption, use cases, journeys, outcomes, unmet needs, and product decisions. Flowlines connects both views without replacing infrastructure monitoring.
Does MCP observability require instrumentation?
Yes. The server needs compatible telemetry. The Flowlines plugin can guide a repository through an OpenTelemetry-based integration after the data boundary is reviewed.
Can Flowlines identify MCP users?
Flowlines can group calls by a person, agent, workspace, account, or MCP client when the server is permitted to report a stable identity. Analytics identity never replaces the server's authorization checks.
Does Flowlines replace Grafana, Datadog, Sentry, or OpenTelemetry?
No. Keep the operational tools you use for infrastructure, errors, logs, and traces. Flowlines adds behavioral and product analysis across users, sessions, use cases, and outcomes.
Can Flowlines notify a team about MCP behavior?
Yes. Configure notifications for recurring failures, unmet product requests, buying intent or custom behavior signals. Route findings to Slack, email or a webhook with real-time delivery after analysis, or daily and weekly digests. Analysis and grouping affect delivery time.
MCP server setup
Connect Flowlines. Ask to onboard.
Your assistant prepares and applies an OpenTelemetry plan, then checks the result. No proprietary SDK or extra tracing platform required.
Your own server. Your own data.
Start where you already work.
one connection, then your use case
Add the MCP connection or the official plugin.
Get a plan for the MCP server in your repository.
Your assistant adds OpenTelemetry instrumentation and runs the tests.
Ask Flowlines to verify the telemetry and fix any gaps.
100 MCP sessions a month, free