> ## Documentation Index
> Fetch the complete documentation index at: https://docs.thecontextcompany.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Insight Search: investigate production in natural language
> Ask questions about runs, tools, models, failures, costs, users, and feedback in plain English and get evidence-backed answers.
Insight Search answers the question: **why is this happening?**
It is a multi-step analytics agent with full SQL access to your telemetry. Ask a question in natural language and it plans queries across runs, sessions, tool calls, model calls, failures, patterns, feedback, cost, and metadata, then returns a written answer with links to every run behind each finding.
Insight Search requires a **Pro** or **Enterprise** plan. Upgrade from [Settings](https://www.thecontextcompany.com/prod/settings).
## Where to use it
* The [dashboard](https://www.thecontextcompany.com/prod/search) search bar
* The [Slack](/integrations/slack) bot (`@The Context Company`) in any subscribed channel
* Your IDE or coding agent through [MCP](/access-data/mcp) (Cursor, Claude Code, and any MCP-compatible client)
* Programmatically through the [REST API](/access-data/api#insight-search)
## What you can ask
Insight Search is not restricted to a fixed set of dimensions. It can join across runs, sessions, tool calls, models, patterns, failures, users, organizations, feedback, latency, and cost.
### Broad questions
**Frustration analysis**
> Find the 3 biggest reasons users were frustrated with the agent in the past week and explain why they're happening.
**Cost trends**
> How have my agent costs changed over the past month? Which models and tools are driving the most spend?
**Weekly health check**
> Give me a summary of how my agent performed this week compared to last week. What got better, what got worse?
**Topic breakdown**
> What are users asking about most? Which topics have the highest failure rate and which have the most positive feedback?
**User satisfaction**
> Are users generally happy with the agent? Show the ratio of positive to negative feedback and what the negative feedback is about.
### Execution-aware questions
These are the questions that transcript-only analytics cannot answer.
**Silent tool failures**
> Find runs where a tool returned an error but the final response looked successful.
**Tool ordering issues**
> Find runs where the search tool was called after the summarize tool for users on the enterprise plan where the final response didn't answer the user's question.
**Multi-turn failures**
> Show me sessions with more than 5 back-and-forth messages where the user eventually gave up. What was the agent getting stuck on?
**Tool error chains**
> Find runs where a tool call failed immediately after another tool succeeded. Which tool pairs fail together most often?
**Regression detection**
> Compare error rates for the `generate_report` tool this week vs. last week. Did a specific model start failing more?
**Slow paths**
> Find runs over 30 seconds where the user waited on more than 3 sequential tool calls. What tools were in the chain?
**Impact by customer**
> Which organizations were most affected by tool errors this week, and how many runs each?
## How answers are structured
Every Insight Search answer includes:
* A written summary of what was found
* The specific runs or sessions behind each finding, with links straight into their [traces](/investigate/traces)
* The relevant metric (count, cost, latency, tokens, feedback ratio, etc.) alongside the evidence
Open any linked run to see the full execution: prompt, response, model steps, tool calls, arguments, results, errors, cost, and metadata.
## Chat history
Your searches are saved locally so you can return to previous investigations. Start a new chat anytime or continue from where you left off. In Slack, follow-up questions in the same thread keep context.
## Insight Search in Slack
Mention the bot (`@The Context Company`) in any subscribed channel. Threaded follow-ups preserve context, so you can go from "what are the most common errors?" to "show me examples of the first one" without repeating yourself. When the bot notices a behavior worth tracking, it can suggest creating a [pattern](/analyze/patterns) directly from the thread.
## Insight Search over MCP
Insight Search is the primary [MCP tool](/access-data/mcp#insight_search). Point Cursor or Claude Code at your data and your coding agent can find the affected runs, pull the offending trace into its context, and propose a code fix in place.
## Limitations
* Queries production data only.
* Answers are grounded in your telemetry. If a dimension is not captured (see [what The Context Company captures](/what-tcc-captures)), Insight Search cannot analyze it.
* Complex cross-dimensional queries may take a few seconds while the agent plans and runs multiple SQL passes.
## Related
Every result links into the underlying trace.
Use Insight Search from your IDE.
Investigate from a Slack thread.
Programmatic access.