Token Optimization and Analytics
Token Optimization and Analytics helps admins understand where token usage comes from and reduce unnecessary cost or context.
Both surfaces live on the same page, Monitor > Tokens, which is titled Token Optimization. It has three views:
| View | What it does |
|---|---|
| Analyze | Shows token usage over a date range: conversation spend, response tokens, init context tokens, tool token usage, toolkit context savings, biggest spenders, and models used. |
| Optimize | Lists recommended actions with estimated monthly savings, plus MCP context, toolkit context, skill size, and tool response opportunities. The Output Format control sits here. |
| Devices | Breaks usage down by the devices reporting to your organization. |
Use Analyze to find the largest sources of token usage. Use Optimize and its recommended actions to reduce them.
How the pieces fit together
AI agents consume tokens in several places:
| Source | Why it matters |
|---|---|
| MCP context | MCP servers expose tool descriptions and input schemas. Large servers can add a lot of initial context before the agent calls any tool. |
| Toolkit context | Toolkits expose curated subsets of tools. Smaller, focused toolkits reduce the initial context shown to agents. |
| Skill context | Skills are loaded into every AI conversation. Oversized skills raise the baseline cost of every session. |
| Tool responses | Tool outputs can be large. Response mappings and output-format settings reduce the number of tokens returned to the model. |
The normal workflow is:
- Enable Token Usage Analytics.
- Review recommendations on the Optimize view of Monitor > Tokens.
- Inspect high-context MCP servers, toolkits, and skills.
- Create smaller toolkits, move long skill content into references, or tighten tool responses.
- Adjust the output format if tool responses are a major source of usage.
Pages in this section
- Enable Token Usage Analytics - turn on the beta analytics feature in Settings.
- Review Token Usage - read the Tokens dashboard and recommended actions.
- MCP Context - find MCP servers with large tool-description and schema footprint.
- Toolkit Context - use toolkits to reduce initial context.
- Skill Token Analysis - identify oversized skills and extract references.
- Tool Response Optimization - find unmapped high-token tool responses.
- Output Format Optimization - set the tool output format from the Optimize view.