How To Save 90% of Claude Code Token Usage

Ways to Reduce Claude Code Token Usage

The video presents four ways to reduce Claude Code token usage, potentially by as much as 90%. Savings depend on the project and workflow, and each method involves trade-offs—especially around accuracy and added complexity.

1. Index the Codebase with CodeGraph

CodeGraph creates a searchable graph of the codebase, allowing Claude Code to find relevant files through semantic queries instead of repeatedly searching and reading files.

The tutorial demonstrates initializing CodeGraph, checking its status, and running a query. Claude Code can also use its CLI commands.

Trade-offs: The index must stay synchronized with the code, or it may return outdated results. Searches may also miss directory-specific context files such as CLAUDE.md. Indexing helps find files, but Claude still needs to read the relevant ones.

2. Compress Command Output with RTK

RTK reduces noisy CLI output—such as long Git logs or test results—before Claude Code processes it. In the demonstration, RTK summarizes results and omits repetitive lines.

Trade-offs: Compression is lossy and may hide a useful warning or log message. The speaker recommends disabling it when debugging requires detailed output.

3. Shorten Responses with Caveman

Caveman applies a session mode—Light, Full, or Ultra—to shorten Claude Code’s responses and reduce the conversational history carried forward.

The speaker demonstrates a project explanation with and without the tool, reporting roughly half as much output in the example.

Trade-offs: Over-compression can remove context from the agent’s feedback loop and lead to less reliable answers. The speaker uses Ultra for routine work but would turn it off when detailed reasoning or planning matters.

4. Manage Sessions and Choose Models Deliberately

• Use /context to inspect what is consuming tokens, including large context files or loaded tools.
• Use /clear when starting an unrelated task and /compact to manage accumulated context.
• Use /model to switch models. The speaker recommends trying Haiku or Sonnet for suitable, repeatable tasks, such as browser navigation or scheduled jobs, while favoring the most reliable model for significant programming, planning, and deep dives.
• Use Plan mode before execution. For UI work, consider designing with tools such as Figma before asking Claude Code to implement the interface.

Overall

These techniques can substantially reduce token use, but they also add configuration and potential failure points. The speaker recommends choosing tools based on the task rather than enabling every compression strategy indiscriminately.

Main speaker/source: John Kim, the video’s presenter.