Skip to content
Halv
GitHub
Download
NEWHalv is now open sourceRead the announcement →

Open source.
Half the spend.

Halv runs Claude Code, Codex, Kimi and GLM in one desktop workspace, on the plans you already pay for. The app and its local savings engine are now open source. Halv Cloud adds the router that sends each task to the model it needs.

Download for macOSFREE★ Star on GitHub
Windows and Linux →
SWE-rebench · 42 paired runs
57.1%
lower cost
per correct
answer.
Halv / pass$5.79
Codex alone$13.50
Same pass count: 25 of 42 in both workflows.
Halv desktop workspace with project folders, agent tabs, Claude Code and Codex side by side, and a savings meter.
Inside Halv · Claude Code and Codex side by sideWorks with your Claude, ChatGPT, Moonshot or Z.ai plan
01 / Two halves

Almost all of Halv is free. The rest pays for itself.

½ — Open source

Halv

$0forever

The whole workspace and every local savings layer. Download it, run it, keep it.

  • Four agents, one window
    Claude Code, Codex, Kimi and GLM, each on your own subscription.
  • Up to 32 live panes
    Native Pro-mode terminals and subagent fan-out.
  • Crux code index
    Callers, references and impact, answered from an index instead of file-by-file reads.
  • RTK + Headroom
    Command output filtered and context compressed, on your machine.
  • Your control
    Local OS sandbox, per-project approvals, turn diff and revert.
  • Live savings meter
    Dollars saved in every conversation, measured as you work.
Download freeView the source
Apache 2.0 · macOS · Windows · Linux
½ — Paid

Halv Cloud

$10/ month

The router that decides how much model each task gets, and the place your savings add up.

  • JEV routingHow it works ↓
    Halv Cloud rates each brief with JEV, with k calibrated on the bench.
  • Savings dashboard
    History across sessions, projects and every machine you work on.
  • Automatic account rotation
    When one plan hits its window, work moves to your next account.
  • Phone relaySOON
    Your sessions, relayed to your phone.
Why this half is paid

It runs on our servers, not on your machine. JEV runs a model for every route_task call, keeping k calibrated means re-running the bench on real models, and your history is stored and synced across machines. The US$10 pays for that. Everything that runs on your computer stays free.

$192.83saved in one benchmark run. At US$10 a month, that is 19 months of Halv Cloud. See the run →
Get Halv CloudHow routing works
Cancel in two clicks · Agent plans stay separate
02 / Halv Cloud · JEV routing

One call decides what each task costs.

The agent you chat with briefs each task and calls route_task. From version 1.0.2, Halv sends a brief only after you say yes to brief sharing. On Halv Cloud, JEV reads the brief and picks the lane, the model and the effort. Easy work goes to cheaper models, hard work to stronger ones, and a task that fails its checks escalates a tier.

  1. 01

    The brief

    Your coordinator agent writes a self-contained brief for each task.

  2. 02

    route_task

    Once you say yes to brief sharing, the call goes to Halv Cloud, where JEV reads the brief with a k calibrated on the bench. A brief can hold code.

  3. 03

    Lane, model, effort

    The worker runs on your Claude or Codex plan. Failed checks escalate to a stronger tier.

halv.cloud/jev· route_task
routed 6/6k · calibrated on SWE-rebench
BriefLaneModelEffortStatus
Rename userId to accountId in 3 filescodexgpt-5.6-sollow✓ routed
Add an idempotency key to the Stripe webhookcodexgpt-5.6-solmedium✓ routed
Fix the race in session refreshclaudeopus-5.5high✓ routed
Update the checkout e2e snapshotscodexgpt-5.6-solminimal✓ routed
↳ checks failed, escalating a tiercodexgpt-5.6-solhigh↑ escalated
Write the ledger table migrationclaudeopus-5.5medium✓ routed

Every routed task shows its lane, model and effort on the task board.

  • Calibrated on the bench.

    The k behind every routing decision is calibrated on our public SWE-rebench runs, the same ones where Halv recorded 57.1% lower cost at the same pass count. Routing runs on our servers, so every recalibration reaches you without an app update.

    Read the benchmark ↗

  • Your plans do the work.

    Routed workers run on the Claude and ChatGPT subscriptions you already pay for. JEV is included in Halv Cloud, with no separate routing fee.

  • Your call, always.

    Tell the coordinator to implement with Codex or Claude and routing stays in that lane. Turn routing off in Settings and you pick the model yourself.

03 / Proof · SWE-rebench interim

Same pass count. 57.1% less cost.

Across 42 paired repetitions, both workflows passed 25 runs. Recorded model cost fell from $337.50 to $144.67 with Halv.

14 tasks × 3 repetitions. Every outcome.Verifier passedDid not pass
Codex alone
25 / 42
Codex + Halv
25 / 42

Tasks from ten open-source repos, including ArcadeDB, nanobot, hugegraph and pulsar. Hover a cell for the task, click to open its record.

MetricCodexCodex + HalvChange
Verifier passes25 / 4225 / 42No change
Recorded model cost$337.50$144.67−57.1%
Cost per correct answer$13.50$5.79−57.1%
Total tokens48.8M170.1M+248.9%

Halv used 3.49× as many tokens. Model routing reduced recorded cost.

Interim result: the first 14 of 111 planned tasks, Astra medium, Codex 0.155.1. Costs exclude invalid infrastructure attempts and a rolled-back policy experiment. The free half has receipts too: Crux answered 48 of 50 code-navigation questions correctly against 33 with grep, at 24% less cost per answer.

Read the benchmark ↗Explore all 42 pairs →Download the evidence ↓Crux benchmark ↗
04 / Compare

Everyone runs agents. Halv makes them cheaper.

We mapped the tools developers use to run Claude Code and Codex today. The first three rows are table stakes. The rest is where the bill comes from.

Capability½ HalvT3 CodeOrcaConductorEmdashSupersetCodex appClaude Code
Table stakes
Open sourceYesYesYesNoYesYesCLI onlyNo
macOS, Windows and LinuxYesYesYesMac onlyYesMac onlyYesYes
Claude and Codex side by side, on your plansYesYesYesYesYesYesOpenAI onlyAnthropic only
The bill
Command-output filtering, built inYesNoNoNoNoNoNoNo
Code index for callers and referencesYesNoNoNoNoNoNoLSP plugins
Routes each task to a lane, model and effortYes CloudManualManualManualManualManualManualManual
Live dollars-saved meterYesNoUsage onlyNoNoNoNoNo
Paired cost benchmark, evidence publicYesNoNoNoNoNoNoNo
Rotates accounts when a plan runs dryYes CloudNoManualNoNoNoNoNo

From each product's public site and docs, October 2026. Codex and Claude Code are compared as used on their own; Halv runs both. Cloud means included in Halv Cloud.

Also on the map
  • Cursor
  • Vibe Kanban
  • Sculptor
  • Warp
  • Parallel Code
  • OpenCode
They help you run more agents. Halv makes each one cost less.
05 / Open source

Read the code that reads your code.

Indexing, compression and command filtering run on your machine, in code you can audit, fork and ship.

  1. 01
    Audit every request.
    The client is open, including the part that talks to Halv Cloud. See exactly what a route_task call sends.
  2. 02
    Keep the free half.
    No trial clock on the workspace or the local engine. It is licensed under Apache 2.0, so it stays yours.
  3. 03
    Send a pull request.
    New agents, new panes, better filters. Issues and pull requests are open.

Get Halv

06 / FAQ

Questions we get asked.

Something else? Write to hello@halv.ai or open an issue on GitHub.

What exactly is open source?

The Halv desktop app and its local engine: the workspace and native terminals, Crux, the RTK and Headroom integration, the sandbox, approvals, turn diff and revert, and the local savings meter. Download it, or build and ship your own from source.

What does Halv Cloud add?

JEV routing and the cloud savings dashboard, with history across every machine you use, plus automatic account rotation. The phone relay joins Halv Cloud when it ships. US$10 a month. There is no trial: the free app is the trial.

If Halv is open source, why is Halv Cloud paid?

Because it costs money to run. Routing runs a model on our servers for every task. Keeping k calibrated means re-running the bench on real models; the interim SWE-rebench run alone recorded $482 in model cost across both arms. History and multi-machine sync need storage. Everything that runs on your computer costs us nothing to give away, so it is free and open source.

Why does routing run on your servers?

Because that is where k lives. route_task needs a k calibrated on the bench, and keeping JEV on Halv Cloud means every recalibration reaches you as soon as it ships, with no app update. Turn routing off in Settings and you pick the model yourself.

Does Halv Cloud see my code?

Indexing, compression and command filtering run on your computer. To route a task, Halv Cloud receives the brief your coordinator agent wrote for it. From version 1.0.2, Halv sends a brief only after you say yes to brief sharing. A brief can hold code, file paths, commands and secrets. Versions 1.0.1 and older do not ask. While you are signed in to Halv, they send each brief that an agent passes to route_task or route_tasks. This also happens when task routing is off or your plan has no Halv Cloud. Your model provider receives the prompts and context it needs, as it does today. The client code that sends all of this is open, so you can check.

Will every task cost 57.1% less?

No. Halv cost more on four of the 14 tasks and used more total tokens. Equal aggregate pass counts hide different solved tasks. This interim Codex result does not establish Claude Code savings or reduce fixed subscription prices.

Which agents and platforms does Halv support?

Claude Code, Codex, Kimi and GLM, each through its native CLI, signed in with the account you already have. macOS on Apple silicon and Intel, Windows x64, and Linux x64 as AppImage or deb.

Download it. Run it. Watch the meter.

The meter starts at zero and counts only what Halv actually saved you. A day of your own work is the whole pitch.

Download for macOS★ Star on GitHub

Free and open source · Halv Cloud US$10/month