Skip to main content
This quickstart takes you past “telemetry is arriving” and through the workflow you will actually use: send a real run, inspect its trace, attach user and session identity, and see The Context Company detect a behavioral signal. Plan on about 10 minutes.
1

Create an account and grab a key

Sign up at thecontextcompany.com and copy an ingestion key (prefixed tcc_prod for production or dev_ for your personal dev environment) from Settings.Set it as TCC_API_KEY in your environment. See Environments for the difference between dev, prod, and local mode.
2

Instrument your agent

Pick your framework and follow its integration page. Each one installs a small package and hooks into the framework’s native tracing.

Vercel AI SDK

Claude Agent SDK

Mastra

LangChain & LangGraph

CrewAI

Agno

If you use another framework, use the Python or TypeScript custom instrumentation, or the OpenTelemetry integration.
3

Send a real interaction, not a hello world

Trigger your agent the way a real user would (a real prompt, real tools, real downstream calls). One good run is enough.On that call, attach the reserved metadata that turns the run into something you can analyze:
Setting tcc.conversational: true enables behavioral pattern analysis on the run. Framework-specific syntax lives on each integration page.
4

Open the run

In the dashboard, open the run you just sent. You will see:
  • The full prompt and response
  • Every model step in order
  • Every tool call with its arguments, result, latency, and status
  • Tokens and cost per step
  • The user, organization, session, and agent it belongs to
This is the trace. It is the primary evidence for every failure, pattern, and Insight Search result later on.
5

Filter by user, org, or agent

From the runs list, filter by the tcc.userId, tcc.orgId, or tcc.agent you just set. These are first-class filters, not custom metadata. See Users and organizations for the analyses this unlocks.
6

Send a few more runs, then check patterns

Send a handful more conversational runs (a mix of good, bad, and ambiguous is ideal).Open Patterns. Three built-in patterns run automatically on conversational runs:
  • Frustration — user expresses annoyance
  • Confusion — user indicates they don’t understand
  • Task failure reported — user reports the agent didn’t complete their request
Open any pattern to see matching runs, the confidence score, and the evidence the classifier used.
7

Ask a question in natural language

Open Insight Search and ask something specific about your data:
What are the most expensive runs in the last day and which tools did they call?
You get a written answer, the runs behind it, and links straight into their traces. Insight Search is also available from Slack and MCP.
8

Optional: connect Slack

Connect Slack to receive pattern alerts, subscribe channels to Recaps, and run Insight Search from a thread. Add the bot with /subscribe in any channel.

​
What you have now

  • Runs, sessions, users, organizations, and agents flowing in with the identity your product knows about
  • Full execution traces (model calls, tool calls, arguments, results, errors, cost, latency) per run
  • Automatic pattern detection on conversational runs
  • Natural-language investigation over the whole dataset from the dashboard, Slack, MCP, and the REST API

​
Next

Walk through a silent tool failure

See exactly how execution-aware analytics catches a bug that transcript-only analytics would miss.

Concepts

Runs, sessions, users, organizations, and the reserved metadata keys.

Patterns

Add custom classifiers for domain-specific behaviors.

Traces

How trace data drives every analysis feature.