Documentation

Connect Flowlines to production telemetry.

Application-agent traces can connect through OpenTelemetry, Langfuse, or LangSmith. MCP servers use a separate skill-guided instrumentation path.

Book a setup meeting →Instrument an MCP server →

Selected path

Point your exporter at Flowlines.

Use OTLP/JSON or protobuf and pass the namespace key in x-flowlines-api-key.

OTEL_EXPORTER_OTLP_TRACES_ENDPOINT=https://api.flowlines.ai/v1/traces

Selected path

Connect one project or add a second exporter.

Keep Langfuse as your trace store. Flowlines adds cross-session analysis.

Mode: read-only polling or OpenTelemetry push

Selected path

Read one project or mirror its runs.

Threads and runs become the production sessions Flowlines analyzes.

Mode: read-only polling or OpenTelemetry export

Selected path

Instrument the server tool boundary.

Review exported fields, add AGNTCY Observe or OpenTelemetry, then verify one journey.

Use the Flowlines MCP observability skill

01 / Choose a path

Connect what you already collect. Instrument what is missing.

AApplication agents

Connect OpenTelemetry, Langfuse, or LangSmith when your traces already carry the context Flowlines needs.

Book a setup meeting →

BMCP servers

Use the Flowlines skill when the server does not yet emit the canonical tool-call span.

Open MCP setup ↓

02 / Existing traces

Keep the tracing stack already running in production.

Connect Langfuse or LangSmith read-only, send standard OpenTelemetry, mirror a provider through OpenTelemetry, or import one public LangSmith trace share.

Flowlines analyzes behavior across sessions. Your existing provider can remain the place where engineers debug individual runs.

03 / OpenTelemetry

Send traces to the standard OTLP endpoint.

# OpenTelemetry exporter variables
OTEL_EXPORTER_OTLP_TRACES_ENDPOINT=https://api.flowlines.ai/v1/traces
OTEL_EXPORTER_OTLP_HEADERS="x-flowlines-api-key=<your-api-key>"
OTEL_EXPORTER_OTLP_PROTOCOL=http/json

Flowlines accepts OTLP/JSON or protobuf, including gzipped payloads. Preserve stable session and user identifiers so cross-session analysis can group production behavior correctly.

04 / Instrument MCP

Emit one canonical tool-call span at the server.

The Flowlines skill inspects the server framework and existing telemetry, shows which arguments, results, and identity fields may be exported, then waits for approval.

01Inspect

Find the framework, tool handlers, request context, and current telemetry.

02Approve

Choose which tool data and identities may leave the server.

03Instrument

Add AGNTCY Observe or OpenTelemetry with the smallest compatible change.

04Verify

Run one journey and confirm its intent, tool path, and outcome.

Open the one-click setup ↓

05 / Verify data

Confirm one complete session before reading trends.

Run a representative agent or MCP journey.
Find the session by user, intent, server, or time.
Confirm the user, parent session, tool names, durations, status, and outcome context.
For MCP, verify the tool path appears under the correct use case.

06 / Read an issue

Move from a recurring pattern to the original sessions.

The Issues view separates behavioral failures from reliability errors, groups related behavior, and shows the affected users, use cases, releases, and sessions.

Pattern→Impact→Sessions

07 / MCP journeys

Follow the path from use case to outcome.

See who uses each MCP server, what they are trying to accomplish, typical tool paths, repeated-call loops, failed sessions, outcome coverage, and the original evidence.

Explore MCP observability →

Common setup questions

Do I need a Flowlines application SDK?

Not when your existing application-agent traces already contain enough session, user, call, and outcome context. Connect OpenTelemetry, Langfuse, or LangSmith instead.

Does MCP observability require instrumentation?

Yes. The MCP server must emit the canonical tool-call span. The Flowlines skill can add it through AGNTCY Observe or standard OpenTelemetry after you review the export boundary.

Does installing the plugin change every project?

No. You choose the MCP server repository. The skill inspects it, proposes the change, and waits for approval before editing code.

Can I keep my current trace provider?

Yes. Flowlines works as the cross-session behavioral analysis layer while your current provider continues to store and debug individual traces.

MCP server setup

Add MCP observability from Claude Code or Codex.

The Flowlines skill reviews the server, asks what may be exported, then adds canonical tool-call telemetry with AGNTCY Observe or OpenTelemetry.

Choose your coding agent

Open the setup where you already work.

Open Claude CodeInstrumentation prompt included Open CodexInstrumentation prompt included

One click opens a guided instrumentation task. You approve every change.

01Review the server

The skill finds the framework, tool handlers, and existing telemetry.

02Set the boundary

You approve which arguments, results, and identities may leave the server.

03Add telemetry

The skill makes the smallest compatible OpenTelemetry integration.

04Verify a journey

Run one tool flow and confirm its intent, calls, and outcome in Flowlines.

Prefer the terminal?Show one-line commands

Claude Codeclaude plugin marketplace add flowlines-ai/plugins && claude plugin install flowlines@flowlines

Codexcodex plugin marketplace add flowlines-ai/plugins && codex plugin add flowlines@flowlines

Then use /flowlines:flowlines-mcp-observability in Claude Code or $flowlines-mcp-observability in Codex. Source and release history are in the official Flowlines plugin repository ↗.