Halv: pricing, features, alternatives

Halv is a desktop workspace for coding agents. The site at halv.ai says you run Claude Code, Codex, Kimi, and GLM side by side, on the subscriptions you already pay for. Halv sits in front of those CLIs with a compression engine, a Crux code index, project sandboxes, and a live savings meter. Source is private; public installers live in pedr0v/halv-releases (created 2026-08-04).

On 20 paired SWE-rebench tasks (gpt-5.6-luna, medium reasoning, Codex 0.152.0), the homepage reports Halv used 51.1% fewer tokens per verified correct answer, with a codebase verifier accepting 10 Halv answers versus 7 from Codex alone. It also says Halv used fewer tokens on 12 of 20 tasks and more on 8. Those are vendor measurements with public run logs and SHA-256 checksums on the site, not an independent audit. Results vary by task and model.

Compare Paseo if you want an open-source control plane across more agents, or Happy if you only need a mobile client for Claude Code.

Key Features

Same CLI, less context waste: login stays in Claude Code, Codex, Kimi, or GLM. Halv compresses duplicate files, stale history, and boilerplate before the request. Requests it cannot safely compress pass through unchanged.
Crux index: callers, references, and blast radius come from the index so the agent does not burn context reading the tree file by file.
Chat or native terminal: split panes, saved sessions, diffs, approvals, and a savings meter in one workspace. Several agents can run at once.
Local by default: the FAQ says the repo stays on device; only model calls leave, the same class of request the CLI already sends. Shell commands run in an OS sandbox and need your approval.
macOS, Windows, Linux: Apple silicon and Intel macOS, Windows x64, Linux AppImage and deb.

Limitation: Halv is not a model and not open source. You still pay Anthropic, OpenAI, Moonshot, or Z.AI. Basic caps monthly savings; Unlimited removes the cap. The 51.1% figure is from 20 vendor tasks, not a guarantee on your repo.

Use Cases

Stretch a Codex or Claude Code plan: same login, fewer tokens on noisy tool output.
Run two agents without two windows: Claude Code on one pane, Codex on another, sessions saved.
Index-first navigation: let Crux answer "who calls this" instead of dumping the tree into context.

Pricing

Plan	Price	Notes
Basic	$2 / month	All features. Monthly savings capped.
Unlimited	$10 / month	All features. No savings cap.

Both plans include a 7-day free trial. No credit card to start. Cancel in two clicks, per the homepage. Agent subscriptions stay billed by their vendors.

Getting Started

Download for macOS, Windows, or Linux from halv.ai or GitHub Releases.
Sign in through the agent CLI you already use. API keys are optional.
Open a project, pick an agent, and watch the savings meter on the first prompts.

First-party resource: halv.ai.

Frequently Asked Questions

Do I need a new API key?

No. Halv uses each agent's own login. Keys work if you prefer them.

Does code leave my machine?

The FAQ says no, except the model request the CLI would send anyway.

Is it only for coding?

No. Chat view is for questions, writing, or research. Coding agents, terminal, and Crux are there when you need them.

Alternatives

Paseo: self-hosted control plane for Claude Code, Codex, Copilot, OpenCode, and Pi.
Orca ADE: desktop ADE with agents in isolated worktrees.
Claude Code: the CLI Halv wraps rather than replaces.

Tips

Track savings in-app for a week before picking Basic vs Unlimited.
Keep approvals on until you trust the sandbox on that repo.
Treat the 51.1% headline as a 20-task vendor study. Recheck on your own workload.

Conclusion

Halv is for people who already pay for Claude Code or Codex and want the same CLIs to waste less context. It is a paid desktop layer, not another model. Try the seven-day trial on one real project before you care about the SWE-rebench poster.