Where agent data shows up
Find which Mintive views each data source fills, so you know what to set up when a view you expect to use is empty. Each source sees a different part of an agent's work: a local hook sees the tool calls on one machine, an inference hook sees each prompt before it reaches Claude, and usage telemetry reports the tokens a session consumed. So a view only fills in once a source that sees its data is connected.
Data sources
| Source | What it sees | Set it up |
|---|---|---|
| Local hooks (Claude Code, Cursor, Codex, GitHub Copilot) | Prompts, file reads, commands, and MCP tool calls on each machine where the hook is installed | Claude Code, Cursor, Codex, GitHub Copilot |
| Claude inference hooks | Each governed prompt across claude.ai, Claude Cowork, and Claude Code, before it reaches the model. The model's response isn't part of the delivery, so there are no token counts or cost | Claude inference hooks |
| Usage telemetry | Token counts and estimated cost for each session and model, reported by the agent | Turn on usage telemetry |
| LLM gateway | Every inference call routed through a gateway policy, with tokens, cost, and latency | LLM gateway |
| Virtual MCP servers | Every tool call made through a Mintive Virtual MCP | Virtual MCP concepts |
Which views each source fills
| Source | Live Activity | Dashboard | Usage | Insights: Agent activity chart, users, requests | Insights: MCP server cards and tool report | Budgets |
|---|---|---|---|---|---|---|
| Local hooks | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
| Claude inference hooks | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ |
| Usage telemetry | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
| LLM gateway | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ |
| Virtual MCP servers | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
Live Activity, Dashboard, and Usage are the tabs under Agent Monitor → Agent Activity. Insights is under Agent Monitor → Insights, and Budgets are set on LLM gateway policies.
Three sources also have their own pages:
- LLM gateway → LLM usage shows gateway spend and request volume, filterable by policy, model, harness, and user. Budgets are managed on each policy's Budgets tab, see Budgets.
- MCP activity shows the tool calls made through your Virtual MCP servers. A Virtual MCP call also fills the Insights MCP server cards when a local hook sees the agent make it.
- Tools → Claude usage shows token consumption and billed cost for your Anthropic organization, read live from Anthropic. Turn it on under Enterprise → Claude usage, and members supply their own Anthropic Admin API key when they open it, which Mintive never stores.
Insights
Insights summarizes agent activity across your organization. It refreshes once a day, with data through the end of the previous day (UTC), so new activity appears the day after it happens.
- Agent activity charts users or requests for each agent, daily or weekly. The weekly view only shows complete weeks (Monday to Sunday), so switch to Daily during the first week of data.
- Growing activity lists MCP servers whose calls in the last 7 days are at least 20% higher than the 7 days before, so it needs two weeks of tool calls. Switch the dropdown to New to see servers first used in the last 30 days.
- Active usage lists MCP servers first used more than 30 days ago that are still in use. Servers first used within the last 30 days appear under New instead, so this card stays empty for the first 30 days of monitoring. Switch the dropdown to Inactive to see servers unused for 30 days.
- Weekly report covers 7 full days: total users, total requests, flagged or blocked tool calls, a tool usage chart, and per-user activity. Pick an earlier period from the report selector.
The MCP server cards, the tool usage chart, and the flagged or blocked count come from tool calls, which only local hooks record. Claude Code's and Cursor's built-in tools, such as file reads and shell commands, are left out of the MCP server cards.
Fill an empty view
- Insights MCP server cards or tool report are empty with only inference hooks connected. Inference hooks see prompts, not tool calls. Install local hooks alongside them: once inference hooks are active, local Claude Code hooks keep monitoring tool calls but stop their own prompt checks, so prompts aren't counted twice.
- Usage shows no tokens or cost. Turn on usage telemetry in your monitored clients. Neither local hooks nor inference hooks report tokens.
- You need to cap spend. Route inference through the LLM gateway and set a budget on the policy. Usage telemetry and Claude usage show spend but don't enforce limits.
- The Agent activity chart looks empty in the first week. Switch it to Daily, because the weekly view waits for a complete week.