conon.ai

conon.ai/Tips

MCP Replies Bill Your Context

Connecting servers is nearly free now — tool descriptions load on demand. The cost moved to responses: whatever a server returns lands in your window whole (warning at 10k tokens, truncation at 25k). A CLI tool lets the agent filter output before it touches context; an MCP server doesn't. Let the agent pick per job — it knows both catalogs.

ProTOOLSClaude Code

Why it works

Tool lazy-loading made CONNECTING servers nearly free — the cost moved to RESPONSES. Whatever a server returns lands in your context window whole: there's a warning around 10k tokens and truncation around 25k, and both exist because real servers really do return novels. A CLI tool differs precisely here — the agent can filter its output before it touches context — which is why the right answer per job is 'let the agent pick.'

How to do it

  1. Keep connected servers to the ones you use — connection is cheap; chatty responses are not.
  2. Watch for the large-response warning in the transcript; it names the offender.
    claude code
    ⚠ tool response: 11,842 tokens (search_records)
   large responses fill the context window —
   consider narrowing the query
    The warning names the offender; narrow the query or switch tools.
  3. Bulk-output jobs (big exports, long logs) → prefer a CLI tool the agent can filter through.
  4. Interactive service calls (create a ticket, fetch a record) → the MCP server is perfect.
  5. When unsure, delegate the choice: 'pick the most token-efficient tool for this.'
  6. Recurring pipelines deserve an explicit choice once — then record it in CLAUDE.md.

Copy this

prompt
This job will produce a lot of output. Choose the most token-efficient route — MCP server or a CLI tool you can filter — and tell me which you picked and why before starting.

What to watch for

  • Truncation at the cap is silent-ish: the response just ends. If results feel incomplete after a big tool call, suspect the cap before suspecting the service.
  • Uninstalling every server to 'save tokens' optimizes the cheap half — an unused connected server costs almost nothing until it's called.

Sources

Original: conon.aiDocumentation: code.claude.com

Read next

No Verification, No Green Light
The docs' hardest rule: if you can't verify an output yourself, don't let the agent proceed until IT has. Make self-checking part of done — run it, screenshot it, list what was checked. For anything client-facing, 'it looked right' has never once been a defense.
Claude Code
Cap Research With A Deliverable
'Look into this' with no boundary is how an agent reads fifty files and stalls — the docs list boundless research among the five classic failures. Scope every investigation to a named artifact: 'a one-page summary naming the three files to change'. The deliverable is the fence.
Claude Code
Tell Compaction What To Keep
When a long conversation gets squeezed, the agent decides what survives — and may drop exactly what you needed. One line in CLAUDE.md fixes it forever: 'when compacting, always preserve the list of changed files and my requirements'. Works in every session, no reminders, and /compact accepts the same guidance inline.
Claude Code
Compact Early, On Your Terms
Auto-compaction fires when the window is nearly full — which is after quality already sagged, and the automatic summary guesses at what mattered. Compact earlier and by hand, or ask the agent to lower its auto-compact threshold, so the squeeze happens while the session is still sharp.
Claude Code

We build and operate the systems behind tips like this one — agents, automations and internal tools, shipped into production. See what the studio does →