MCP tool definitions can eat 190K tokens: unloading one server cut context from 200K to 10K
shoumikgoswami · reddit · 2026-10-29
A Reddit post warns that an agent's biggest prompt cost is often the MCP tool definitions loaded before the first message. In a DeepSeek harness repo discussion, one user reported resident context dropping from 200K to 10K tokens after unmounting one large MCP server (single setup, not a controlled benchmark).
Practical checklist
- Check harness docs/logs: does every request carry the full tool list?
- Compare prompt-token usage with and without servers, keeping everything else constant.
- Find the biggest contributors — verbose usage instructions inside tool descriptions made some schemas huge.
- Disable unneeded servers or allow-list specific tools.
- Verify deferred tool loading actually works with your provider.
Measure the baseline before spending hours shortening prompts.
More from coding & agent
- Evals: The Most Underrated Skill in AI Engineering, From Error Analysis to Versioned Golden Datasets — sudoraohacker · 2026-10-29
- Deep dive: editing videos with Claude Code and Opus 5.5, from prompts to a 31K-view short — aakashgupta · 2026-10-29
- This creator spent 40 hours turning Claude into a video editor — and published the full free harness — aakashgupta · 2026-10-29
- Reading PRs through behavior tests: green lines mean additions, red lines mean changes — samgoodwin89 · 2026-10-29
- spawn: 10-person company spends 30x payroll on AI tokens, ships like Anthropic — Scobleizer · 2026-10-29
- SQL hidden in a string literal bypasses LLM-to-SQL gates; schemagate 1.15.1 fixes it — Temporary_Papaya1058 · 2026-10-29