Halv
Halv is a local-only desktop application that reduces token usage in AI coding sessions by compressing the context your coding agent sends to the underlying model. It works with existing subscription-based agents such as Claude Code, Codex, Kimi, and GLM, intercepting and rewriting requests on your machine before they leave. According to Halv, this can cut usage roughly in half while maintaining the same model and answer quality. The app is available for macOS, Windows, and Linux, and is free to test with no credit card or API key required. Halv achieves these savings through three local systems: compression, a Crux code index, and command filtering. Compression removes duplicated files, stale history, and repeated instructions from conversation context. The Crux code index answers structural questions about a codebase — callers, references, and impact — without the agent reading whole files one by one. Command filtering trims noisy command output, such as test logs, dependency trees, and install chatter, down to only the content that matters. Each of these systems is measured separately and visible in a live token meter. Because all processing happens on the user's machine, nothing leaves the computer until the optimized request is sent to the model. The tool provides per-request 'receipts' showing what content was kept, what was cut, and how many tokens were saved. If a request cannot be compressed safely, Halv lets it pass through untouched, preserving correctness. The emphasis on local processing also meets privacy requirements for developers using cloud coding agents with proprietary code.
- Category
- Code
- Pricing
- Free
Problem solved
Coding agents like Claude Code and Codex resend the entire conversation history and full file contents on every turn, causing token budgets and subscription quotas to drain quickly. Halv solves this by rewriting and compressing those requests locally before they are sent to the model, reducing token usage by about half without changing the model, subscription, or answer.
Use cases
- Reducing token consumption in long Claude Code sessions
- Extending the usable time of a Codex subscription window
- Lowering costs when using Kimi or GLM coding agents
- Filtering verbose test logs and command output before model processing
- Indexing code structure to avoid re-reading entire files
- Maintaining privacy and code confidentiality while using cloud models
- Monitoring token savings live on a per-request basis
Audience
- Developers
- Startups
- Hobbyists
- Enterprise
Pros
- Reduces context size and token usage by roughly half
- Works across multiple popular coding agents
- Local-only processing keeps code on the machine
- No API key or credit card required to try
- Detailed per-request token savings and receipts
Cons
- Requires installing and running a desktop application
- Supported only for specific coding agents
- Potential for reduced accuracy if compression logic fails
- Desktop app consumes local system resources
- Not a cloud service; may not suit all workflows
Official website: https://halv.ai/
Canonical LambdaBase page: https://www.lambdabase.com/tools/halv