Windows desktop app

AI Hive

I built it when one terminal per assistant stopped scaling, somewhere around the fifth project. It lets me keep several workstreams moving at once and see all of them in one place.

Source on GitHub Changelog

AI Hive5h 42%

Agent 1HiraKataQuestopus

make the stroke-order demo replayable from the guide panel
Read components/CharacterGuidePanel.kt
Read animation/StrokeAnimationEngine.kt
The engine already has reset(). Adding a replay button that calls it and restarts the clock.
Edit CharacterGuidePanel.kt  +24 -3
Bash  ./gradlew assembleDebug
BUILD SUCCESSFUL
install it on the Pixel
Bash  adb install -r app-debug.apk
Success

Agent 2HiraKataQuestsonnet

review the scoring change, look for edge cases
Read scoring/GeometryScorer.kt
! a hooked stroke scores 0 if the pen lifts early
Edit GeometryScorer.kt  +11 -2
Bash  ./gradlew test --tests *GeometryScorer*
GeometryScorerTest passed
Posted to the board: "scorer fix, only touches GeometryScorer.kt"

Agent 3HiraKataQuestopus

typing "o" for を in the quiz gets marked wrong
Read model/RomajiMatch.kt
The deck spells を "wo" but it's pronounced "o". Accept "o" for を only, never "wo" for お.
Edit RomajiMatch.kt  +6 -1
Edit RomajiMatchTest.kt  +18
Bash  ./gradlew test --tests *RomajiMatch*
RomajiMatchTest passed

Agent 4HiraKataQuestsonnet

a clean エ scores 65 with "Looks more like 工"
Read .aihive/board.md
Agent 2 is in GeometryScorer.kt. Staying out of it.
Read scoring/KanaLookalikes.kt
ML Kit can't tell katakana エ from the kanji 工, they're the same drawing. A twin match should count as recognized.
Edit KanaLookalikes.kt  +14 -1
Bash  ./gradlew test --tests *KanaLookalikes*
KanaLookalikesTest passed

A mock-up of four agents working on HiraKataQuest at once. They check the shared board before touching a file.

What it does

Many agents, real terminals, one window. Each tile is a Windows pseudo-console (ConPTY) running Claude Code, Gemini, Grok or a plain shell, so every CLI behaves exactly as it does on its own. Grids go from 1×1 to 4×3.
Agents that don't trip over each other. Agents in a workspace share an activity board through an MCP server. Before one edits a file, it can see that another agent already claimed it.
Work keeps running when you look away. Each workspace maps to a project folder and keeps executing while hidden. Its sidebar badge stays green while agents are healthy and pulses amber while one is working.
Waits out usage limits for you. It reads plan usage from the CLI into the top bar. When a limit runs out it can wait for the reset and resume the agents on its own.
See what each agent is doing at a glance. Every agent row shows its task, or Claude Code's own conversation title read live from the transcript, so you don't have to read its terminal.
Per-agent model, effort and mode. Dropdowns map to the real CLI flags. The card header follows whatever you change inside the terminal, and the permission mode survives a restart.
A sidebar you can organize. Drag agents and workspaces to reorder them, and group workspaces into collapsible categories. The layout persists across restarts.

How it was built

I started AI Hive on 5 July 2026, when I was running more coding agents than I could follow in separate terminal windows. Since then it has been built by the agents it runs: they take tasks, coordinate on the shared board and ship through pull requests, and every release gets a numbered changelog entry.

Decisions that mattered

Session data is never only in memory. Adding or removing an agent saves at once, saves are atomic (a temp file, then a rename) and a failed load keeps the old file as a backup instead of deleting it. Each rule has a regression test, and the README traces several of them to sessions where agents were lost.
One copy at a time, enforced by Windows. A named mutex refuses a second launch, so two copies can never overwrite each other's session file.
Closing a card kills everything it started. Each card's processes run in a Windows Job Object, so a shell's grandchildren die with it, even if the app itself crashes.
Agents work in lanes. An agent can get its own git worktree. It flags finished work with a "Task done" commit, and an integrator agent merges a batch through one pull request once a GPT-6-Luna review comes back clean.

How it's tested

A smoke suite of 4,454 checks drives the real app headlessly, with real child processes, in up to six parallel workers. One end-to-end test talks to a live Claude Code agent, closes the app, reopens it and checks that the same conversation comes back on the same card, which is exactly where data used to get lost.

Where it stands

In daily use. I run my projects from it, this site included.

Made with

Python
PySide6 / Qt
Windows ConPTY
MCP
pytest

Latest changes

Full changelog

Nov 2, 2026

Version 0.56.1

Internal: the Claude model picker's labels and the per-model effort lookup read the model table through one function, and the Antigravity settings use the same cached reader as Claude's. No change to the app.
Nov 2, 2026

Version 0.56.0

Options > Automation > Keep the PC awake (off by default) stops Windows from going to sleep on its idle timer while AI Hive is open, so agents keep working overnight. The screen still turns off and locks.
Oct 30, 2026

Version 0.55.1

The Options panel lays its settings out in two columns, so the whole panel fits on a 1080p laptop screen. A window too short for it scrolls the panel, and one too narrow for two columns stacks them.