Questions / MCP servers
Straight answers on MCP servers in production.
The questions teams ask once their MCP server is live: who uses it, what breaks, how to get alerted and how to debug sign-in. Earlier application-agent questions follow.
start with the question, not the dashboard
Is anyone actually using my MCP server?
Count distinct users, sessions and returning users in a fixed window, excluding your own team and test scripts, instead of counting tool calls. Flowlines shows who uses your MCP server, from which AI client, which tools they call and whether they come back.
How do I get alerts when errors spike on my MCP server?
Alert on the error ratio per server and tool over a rolling window, with a minimum number of calls and a persistence condition, for example more than 5% errors over 5 minutes with at least 100 calls, sustained for 5 minutes. Add Flowlines to learn which users, clients and use cases the spike affects and to notify the right team after analysis.
How do I track latency for MCP tool calls?
Record the duration of every tool call as a histogram and compare p50, p95 and p99 per server and tool over the same window. Averages hide the slow tail, and percentiles from different instances must not be averaged together. Flowlines shows p50, p95 and p99 latency per tool next to the users and AI clients affected.
How do I monitor several MCP servers in one place?
Give each MCP server its own service name, environment and version in its telemetry, send all of them to the same destination, and compare tools within one server and environment before aggregating across servers. Flowlines organizes observed calls by server, so each one keeps its own users, tools and sessions. Flowlines puts every MCP server in one workspace, with users, AI clients, tools and alerts per server.
How do I know if a change to an MCP tool helped?
Compare the same measures for equal windows before and after the change, with comparable traffic: users of the tool, call success rate, failure classes, repeated calls and how sessions ended. Mark the release with a version, then inspect affected sessions, because a before and after comparison is an investigation, not a controlled experiment. Flowlines compares usage, success and outcomes for the same tool before and after a change.
Can I monitor an MCP server without storing payloads?
Yes. Call counts, duration, error classes, clients, pseudonymous users and sessions all work without storing tool arguments or results. Detailed use cases, buying signals and some outcome checks need additional approved context, so report what you cannot see as unknown. Flowlines supports excluding payloads and still shows usage by tool, AI client and pseudonymous user.
How do I debug MCP OAuth failures?
Walk the MCP authorization flow in order: the 401 response and its WWW-Authenticate header, the protected resource metadata, the authorization server metadata, client registration, the authorization code exchange with PKCE, and audience validation on your server. Most failures come from a mismatched resource URL, a rejected redirect URI or a token issued for the wrong audience.
How do I connect Flowlines to my agent?
Connect Flowlines to Claude Code or Codex with the Use with button. Then choose what to monitor: your MCP server, an application agent, or your coding tool. Each has its own setup and data boundary.
Do I need to install an SDK?
It depends on the data path. Existing application-agent traces can be connected without a proprietary Flowlines SDK. MCP server observability requires an OpenTelemetry-compatible server integration.
How long until I see signal?
A session can appear after Flowlines accepts or imports its trace data. Recurring issue rates, cohort comparisons and trend deltas become reliable only after the selected scope contains enough analyzed sessions, and the product labels insufficient samples instead of presenting a misleading rate.
How do I know if my AI agent is working in production?
Flowlines answers this with analyzed chats, active users, issue-free rate, negative feedback, recurring intents, issues, and sessions that need attention. Open an issue to review the affected users, use cases, and production sessions.
How do I detect agent drift over time?
Agent drift is a recurring change in behavior that becomes visible across sessions, users, intents, or tool journeys. Flowlines groups drift-related issues and keeps the exact sessions attached, so teams can distinguish a pattern from a one-off anomaly.
How do I find the failure modes my logs don't show?
The failures that matter most are often behavioral, not exceptions: a successful response can still be wrong, looping, abandoned, or off-policy. Flowlines separates behavior issues from reliability errors and opens the affected production sessions.
What AI agent issues does Flowlines surface?
Flowlines separates behavior issues from reliability errors. Behavior categories include looping and stuck behavior, drifting, user frustration, abandonment, result quality, hallucination and trust, and safety and policy.
What do my users actually want from my agent?
Users describe what they want in conversations, even when they never submit formal feedback. Flowlines groups analyzed sessions by recurring intent and connects those intents to users, issues, and outcomes.
How do I tell when users are frustrated with my agent?
Frustration can appear as repeated corrections, rephrased requests, negative feedback, or abandonment. Flowlines surfaces user frustration as a behavior issue and opens the affected users and production sessions.
How do I measure user satisfaction without CSAT surveys?
Behavioral analysis does not replace a CSAT survey, but it fills in what surveys miss. Flowlines combines negative feedback, issue-free sessions, repeat usage, activity depth, and affected user groups across analyzed production sessions.
How do I segment my agent's users by behavior?
Flowlines segments observed users with activity and issue evidence, then lets teams create groups using customer, model, environment, or another useful trace property. The Users view shows repeat usage, activation depth, issue coverage, cost, and execution time without inventing a black-box churn score.
How do I find feature requests buried in conversations?
Feature requests often appear inside ordinary agent conversations. Flowlines gives product teams recurring intent groups and the underlying sessions, which provides a practical place to find unmet needs without treating the product as an automatic feature-request tracker.
Why isn't running evals enough?
Evals score the criteria and populations you define. Offline and online evals are essential, but they can still miss recurring production behavior outside those criteria. Cross-session analysis adds changing intents, affected users, drift, and patterns that only emerge at scale.
Is my new prompt working better than the old one?
Compare real production sessions that carry a prompt or release identifier. Flowlines Groups can compare sessions by model, environment, customer, or another useful trace property, but the current product does not invent an automatic causal deployment score.
How do I know what to fix next?
Start with the issue affecting the most important users or sessions, then verify it against the attached evidence. Flowlines gives each issue an impact context, an explanation, a next step, technical detection details, and the exact sessions needed to decide whether the fix belongs in the prompt, tool path, retry policy, or product.
100 MCP sessions a month, free