MCP tool's responses were 62% wasted tokens — here's the fix that cut them by 3x
S_B_B_B_K · reddit · 2026-10-19
A developer measured his MCP brainstorming tool and found 62% of every response was wasted: he had packed the whole idea map into URLs as compressed base64, which tokenizes terribly (1.5 chars/token vs 3.9 for JSON), costing the agent 1,500 tokens to read a URL it can't use.
Fix: store the map server-side and return a short link. Same request before/after:
- 5 ideas: 1,300 → 500 tokens
- 5 ideas + 9 sub-ideas: 2,500 → 1,000 tokens
The real savings are in exploration: 7 rounds of "go deeper on #3" in chat costs 19,000 input + 4,900 output tokens on the main model, versus 1,000 + 50 when clicking the map, where only the path goes to a small cheap model on his side. Lesson: audit what's actually in your tool responses — IDs, URLs, base64 blobs and echoed schemas are easy to miss and expensive on every call. The MCP endpoint (stormap.ai/mcp) is live with a free tier and $0.02/call via x402.
More from coding & agent
- Local LLMs on a 96GB M3 Ultra: Qwen3.8-27B Q8_0 Wins for Coding, With Full Config Shared — FanDiscombobulated38 · 2026-10-20
- Building a Custom CRM/ERP With AI Coding Agents: How Practical Is It? — Possible-Essay-2334 · 2026-10-20
- Multiplayer agent towns could enable fully autonomous agent teams — omooretweets · 2026-10-20
- Vercel Shop goes GA: Vercel and Shopify make AI agents first-class shoppers — evilrabbit_ · 2026-10-20
- Gemini team builds playable 3D Wizard Chess from scratch with multi-agent Blender workflow — GeminiApp · 2026-10-20
- Engineer: my real AI workflow is 10 rounds of dialogue with Claude, not agent factories — brandon_galang · 2026-10-20