Hive Desktop
Download for Windows

Building Hive

How Hive Desktop is built.

Hive Desktop 1.0.0 was built in five weeks by AI coding agents, working in Hive itself. Every change is a card: one model builds it, another reviews it, and Hive’s merge queue merges it once its checks pass. One person directs the work.

Built with Claude Opus · Reviewed by Codex

In numbers

Hive 1.0.0, from its first commit.

As of commit 54892d949 (2 November 2026). How these were counted.

  • 5,298commits3,333 without the merges, from 29 September 2026 to 2 November 2026
  • 2,146cards2,019 of them Done: every change started as one
  • 19releases0.1.0 to 1.0.0
  • 256,000lines of app codeTypeScript and CSS
  • 281,000lines of testsmore than the app itself
  • 8,098unit testsin 611 test files
  • 371end-to-end suites31 of them start the real Claude Code or Codex; the rest, Hive’s fakes
  • 8languagesEnglish and seven translations
  • 1,511agent sessionsabout half in Claude Code, half in Codex
  • 20,697promptsand 325 compactions
  • 68.2 Btokens98% of them cache reads
  • US$17,500at API pricesabout that, for the same tokens; the work ran on subscriptions

The shape of the code

Tests make up nearly half of it.

Every text file in Hive 1.0.0 as a hexagon, sized by its lines and coloured by its folder: the app’s main process, its window and the code they share, then the tests, docs and translations. Brighter files changed more recently. It builds itself in the order the files were first committed.

A honeycomb of 2,000-odd hexagons, one per file of Hive 1.0.0: the tests, in purple, fill nearly half of it, beside the app’s main process, window and shared code, with a legend of each folder’s lines A honeycomb of 2,000-odd hexagons, one per file of Hive 1.0.0: the tests fill nearly half of it, beside the app’s main process, window and shared code, with a legend of each folder’s lines
Drawn from the git history alone, at the same commit as the figures. Open it on its own to see each file’s path and lines.

How the work is done

Every change is a card. Nothing merges until its checks pass.

  • Every change is a card

    A card says what to do and how to tell it’s done. The decisions on it are the user’s, recorded in their words, and the agent checks its work against them before it submits.

  • One model builds, another reviews

    Builders run Claude Code on Claude Opus; reviewers run Codex. A reviewer that finds a problem fails the card with the finding, and the builder fixes it in another round. Since late October, 33% of reviews have sent the work back.

  • Nothing merges until its checks pass

    Hive’s merge queue merges an approved card only once its typecheck, lint and focused tests pass on the merge. The full end-to-end set runs on the integration branch at checkpoints, and for each release candidate.

  • Design before code

    For a bigger change, the builder posts a design on the card first, and the reviewer checks it before anything is built.

  • What wasn’t checked is said

    Each submission carries a run record: which checks ran, on which commit, with what result, and what didn’t run. “Couldn’t do it within the rules” is an accepted outcome, reported on the card rather than worked around.

  • One person decides

    One person directs the work: writes or approves the cards, settles the design questions and tests what needs a human. The agents work within the project’s rules.

The team at 1.0.0: six builders, a pool of three reviewers and one generalist for bug hunts, each one an agent in Hive. Each card submitted for review goes to the next free reviewer.

How these were counted

  • From git, at commit 54892d949, by the site’s build script: commits in its history (with and without merges); releases, the v* tags in that history and 1.0.0; lines in the text files under src (.ts, .tsx and .css) and tests; unit tests, each it() or test() in Vitest’s test files; end-to-end suites, the list the end-to-end runner runs; languages, the folders under locales.
  • From Hive’s own records of the hive project, all time (28 Sep to 2 Nov 2026): cards on its board, archived ones included; agent sessions, prompts and compactions; tokens (input 261 M, cache writes 329 M, cache reads 67.5 B, output 131 M, over 234,061 model requests).
  • At API prices: Hive’s estimate of what the same tokens would cost on Anthropic’s and OpenAI’s public price lists. It isn’t what was paid.
  • Reviews: 1,183 verdicts (792 passed, 391 failed) since late October 2026 (around rc.2), when Hive began recording verdicts, so not all time.
  • The portrait: one hexagon per UTF-8 text file at the same commit, its area proportional to its lines (clamped at 20 to 4,000), without dependencies, build output, lockfiles or the translations other than English.

The same tools are in Hive.

The board, the builder and reviewer, the checks before review and the merge queue that built Hive are the ones it gives you, whether you work with one agent or a team.

Download Hive Desktop for Windows From one agent to a team