The Hidden 30K Tokens Claude Code Uses Before You Type
Read the Portuguese translation →
Author: Leon van Zyl · Watch on YouTube
Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.
Original language: English
Original video: The Hidden 30K Tokens Claude Code Uses Before You Type · Channel: Leon van Zyl · 00:30 · Watch on YouTube
Topics: AI agents, Claude Code, Local and open-source models
Summary
Section titled “Summary”Claude Code uses more than 30,000 tokens of context before the user sends a message, which can be too much for local models with limited context windows. The speaker created a smaller harness for open-weight models running on personal hardware. It removes features those models do not need, such as MCP tools and sub-agents, reducing the required context to 1,400 tokens.
Key points
Section titled “Key points”- Claude Code uses over 30,000 tokens before receiving a user message.
- Local models often have small context windows, particularly on consumer-grade hardware.
- A 32,000- or 64,000-token context window may not have room for Claude Code’s harness.
- The speaker created a small harness for open-weight models.
- The harness strips away components open-weight models do not need.
- The removed components include MCP tools and sub-agents.
- The model’s remaining requirements, including tool definitions and the system prompt, take 1,400 tokens.
Chapters
Section titled “Chapters”| Start | Chapter | Summary |
|---|---|---|
| 00:00 | Claude Code’s context usage | Claude Code consumes over 30,000 tokens before the user sends a message, posing a challenge for local models with limited context windows. |
| 00:30 | A smaller harness for open-weight models | The speaker describes a harness that removes unneeded features and reduces the required context to 1,400 tokens. |
Quotes
Section titled “Quotes”it’s already taking up over 30,000 tokens of my context window. — 00:00
you cannot afford to inject Claude codes 30,000 tokens — 00:00
All the tool definitions, the system prompt, everything this model needs only takes 1,400 tokens. — 00:30
Concepts and entities
Section titled “Concepts and entities”Claude Code, Harness, open-weight models, MCP, subagent, context window
Full transcript (original language)
Section titled “Full transcript (original language)”00:00 Before I’ve sent a single message to Claude, it’s already taking up over 30,000 tokens of my context window. Typically, your local models have a very tiny context window, especially on consumer-grade hardware. If your entire context window is only like 32,000 tokens or 64,000 tokens, you cannot afford to inject Claude codes 30,000 tokens, which is just purely harness stuff, into the context window. So, I decided to create this small coder
00:30 harness, which strips away anything an open-weight model will never need. Open-weight models that run on your own hardware will never need to run things like MCP tools or sub-agents. All the tool definitions, the system prompt, everything this model needs only takes 1,400 tokens.