Q: Can I combine a lightweight startup index with on-demand CLI documentation so the agent only loads what it needs?

Yes, and that progressive disclosure pattern is how the cto-legends router eliminates startup bloat. Instead of injecting 19 full tool schemas at Turn 0, the agent loads a single ~650-token capability index listing module names and trigger verbs. When a task requires a specific tool, the agent runs cto-legends handoff to load the ~800-token CLI recipe for that turn alone.

Question 1

How many tokens do background MCP servers actually consume in Claude Code or Cursor at startup?

Accepted Answer

Each connected MCP server registers its full JSON Schema tool definitions into the system context at session initialization. A single server exposing 12 to 20 tools typically consumes 1,800 to 3,200 tokens depending on nested input schemas. Enabling 15 to 19 background MCP servers burns 30,000 to 42,000 tokens before Turn 1, and that schema payload is re-billed on every turn and tool round-trip.

Question 2

When should I still build or keep an MCP server instead of replacing it with a CLI command?

Accepted Answer

Keep or build an MCP server when your workflow crosses one of three boundaries: first, when the agent runs inside a hosted cloud or browser sandbox lacking a local POSIX shell; second, when the tool must maintain persistent in-memory state across turns, such as a live Playwright browser DOM session; and third, when a multi-tenant platform must broker delegated OAuth tokens on behalf of remote users.

Question 3

How do I stop an AI agent from hanging forever when a CLI tool asks an interactive [y/N] question?

Accepted Answer

Enforce a three-layer non-interactive shell wrapper inside your skill contract. First, export CI=true, PAGER=cat, GIT_PAGER=cat, and NO_COLOR=1 to disable interactive prompts and pagers. Second, pass explicit flags such as --yes and --non-interactive. Third, redirect standard input from /dev/null (< /dev/null) so any unexpected prompt receives an immediate EOF instead of blocking the worker subshell at 2:00 AM.

Question 4

Can I combine a lightweight startup index with on-demand CLI documentation so the agent only loads what it needs?

Accepted Answer

Yes, and that progressive disclosure pattern is how the cto-legends router eliminates startup bloat. Instead of injecting 19 full tool schemas at Turn 0, the agent loads a single ~650-token capability index listing module names and trigger verbs. When a task requires a specific tool, the agent runs cto-legends handoff <module> to load the ~800-token CLI recipe for that turn alone.