Solving Agent Skill Tax: deskill loads MCP skills on-demand, boosting performance
Nilotpal_kakashi · reddit · 2026-08-20
To address the "skill tax" in Agent frameworks (like MCP) where installing too many skills bloats context and degrades performance, the author developed deskill, an MCP server that fetches skill descriptions from GitHub only when actually needed.
Background:
Current systems inject descriptions of all installed skills into every message context. 100 skills consume 3k tokens, costing heavily in long sessions and potentially confusing the model's skill selection.
Solution & Findings:
- Token Optimization: On large models like Opus, on-demand loading matches pre-installed performance, primarily saving costs.
- Performance Boost: On smaller/cheaper models, point-of-use delivery significantly improves task success rates. Smaller models struggle to pick the right tool when staring at a wall of descriptions.
Benchmark:
The author tested 8 models with up to 100 skills and 100k tokens of padding. Results show this is not just a cost-saving trick but also improves smaller model performance. The benchmark is reproducible, and the author invites community contributions.
More from coding & agent
- Devs Stop Reading AI Output, Rely on Goal Verification Mechanisms — BLUECOW009 · 2026-08-20
- 2.5k-star GitHub repo turns classic programming books into rules for AI coding agents — tom_doerr · 2026-08-20
- Gradient Labs launches Collaborate: edit AI agents like code — neal_lathia · 2026-08-20
- Building a Native macOS AI Coding App Using the App Itself — talkaboutdesign · 2026-08-20
- Vercel Agent joins Slack with full production context to execute actions — cramforce · 2026-08-20
- Agent builds OCR app on the fly after read tool removal — mitsuhiko · 2026-08-20