Halv is open source. The desktop app and its local savings engine now use the Apache 2.0 license, and the source is on GitHub.
You can download the desktop app and use it with no trial clock. Halv Cloud, the part that runs on our servers, is the only paid part. It costs US$10 a month, and Halv charges only for what costs money to run.
The first open source release is Halv 1.0.0. The release notes list every change.
What is open source
These parts of the desktop app are open source under Apache 2.0:
- The workspace. It runs four agents in one window, with up to 32 live panes and native terminals.
- The RTK and Headroom integration. RTK filters command output. The local Headroom proxy compresses context on your machine.
- The sandbox and approvals. They give you a local OS sandbox and per-project approvals.
- Turn diff and revert. You can review each agent turn and undo it.
- The local savings meter. It shows the dollars saved in each conversation.
The desktop app also includes Crux, the code index. Crux keeps its MIT license, and you can run it on its own with any MCP-capable agent.
The desktop app has no trial clock. It does not expire, and no local feature stops working after a set number of days. Every local feature works without a Halv account. Halv never blocks chat or terminals. You can run the desktop app, read it, change it, and ship your own build from the source.
Why we did it
Transparency. The code that reads your code is now readable. You can check what the desktop app indexes, what it filters, and what it sends.
Your agents send their requests to their own model providers through the local proxy. Halv sends data to these places only:
- Sign-in and billing. The app talks to the account service at halv.ai when you sign in.
- Savings reports. A report holds token counts and dollar amounts. It holds no prompts, code, or file contents.
- Halv Cloud routing. A routing request sends the task brief to halv.ai.
- Usage events. Halv sends these only if you opt in. Each event carries a random install ID and the app version.
- GitHub. The app checks GitHub for app updates.
- The relay at relay.halv.ai. The app uses it only when you connect a phone. The messages are end-to-end encrypted.
Usage events are opt-in. Halv asks once and lists every event it can send. Halv sends nothing until you say yes, and you can change your answer in Settings. Halv also downloads the newest price table from halv.ai once a day. That request sends no user data. Halv has no other telemetry.
Trust.
Halv indexes your repository and filters what your agents read.
That position needs proof, not a promise.
You can audit every request, including what a route_task call sends.
Halv Cloud receives the brief that your coordinator agent wrote for the task.
The client code that sends this brief is open.
Contributions. Open source lets other people add what we have not built. Examples are new agents, new panes, and new output filters. Halv supports four agents today: Claude Code, Codex, Kimi, and GLM.
Cost. Everything that runs on your computer costs Halv nothing to give away. Indexing, compression, filtering, the sandbox, and the savings meter all run locally. We have no reason to charge for them.
What Halv Cloud is and why it is paid
Halv Cloud is the part of Halv that runs on our servers. It costs US$10 a month and includes four things:
- JEV routing. The
route_tasktool runs on Halv Cloud. It selects the agent, model, and reasoning effort for each task. - A savings dashboard. It keeps history across sessions, projects, and every machine you use.
- Automatic account rotation. When one plan reaches its window, work moves to your next account.
- Phone relay. Your sessions reach your phone. This joins Halv Cloud when it ships.
JEV routing depends on a parameter called k. We calibrate k on our public SWE-rebench runs. Because JEV runs on Halv Cloud, each new calibration reaches you with no app update.
Halv Cloud is paid because it runs on our servers, and servers cost money. Three costs set the price:
- A model call for every route. JEV runs a model for every
route_taskcall. - Calibration. We keep k calibrated by re-running the benchmark on real models. The interim SWE-rebench run recorded $482 in model cost across both arms. That figure covers only the selected runs. It excludes invalid infrastructure attempts and a rolled-back policy experiment.
- Storage and sync. Your history lives on our servers, so it follows you across machines.
Halv Cloud has no trial because the free desktop app is the trial. You can use every local feature before you pay. You can cancel Halv Cloud in two clicks.
What the benchmark shows
In that interim run, we tested 14 tasks three times each. That produced 42 paired runs. Halv and vanilla Codex each passed 25 of them.
- Recorded model cost: $337.50 for vanilla Codex and $144.67 with Halv. That is $192.83, or 57.1%, less.
- Cost per correct answer: $13.50 for vanilla Codex and $5.79 with Halv.
- Tokens: Halv used 3.49 times as many tokens across its coordinator and workers.
- Tasks: Halv cost more on four of the 14 tasks. Equal pass counts do not mean that both workflows solved the same tasks.
This is an interim result for the first 14 of 111 planned tasks. It uses only Codex, so it does not show savings for Claude Code. It does not promise the same savings on your tasks. It also does not lower the fixed price of any agent subscription. Read the short summary, the technical report, or the public evidence.
What does not change
- Your agent plans. Your Claude, ChatGPT, Moonshot, and Z.ai plans stay yours. The agents run on them, and Halv Cloud does not replace them.
- Routing is optional. You can turn routing off in Settings and choose the model yourself.
- Local work stays local. Indexing, compression, and command filtering run on your computer. Your model provider receives the prompts and context it needs, as it does today.
- The client is open. The code that talks to Halv Cloud is open source, so you can read it.
What happens to existing accounts
Existing paid, trial, and lifetime accounts keep everything they had. The desktop app is now free for everyone.
How to contribute
The source is at github.com/halv-ai/halv. Open an issue on GitHub to report a bug or propose a change. We welcome pull requests, in particular for new agents, panes, and output filters.
For a question that does not fit an issue, write to [email protected]. Read the code, run the desktop app, and tell us what you find.