OpenAI Is Buying the Git Host Lovable and Bolt Use. Check What Lives Only There

Posted on Nov 4 • Edited on Nov 4 • AI-assisted

If you have reviewed a change in the Codex app, you have already run code from a company OpenAI is now buying. On November 4, OpenAI said it will acquire Pierre Computer Company, whose open-source Trees and Diffs libraries handle file browsing and diff review inside Codex ( OpenAI, StartupHub.ai, Investing.com ).

No price was disclosed, the deal still needs regulatory approval, and until it closes, OpenAI and Pierre operate as separate companies ( Investing.com ).

The diff viewer is the part most developers have touched. The part to look at if you build on it is Pierre's other product, Code Storage: a Git host for code that agents write.

The open-source side first. Diffs ( @pierre/diffs on npm) renders diffs and code, with split and stacked layouts, line annotations and an edit mode. Trees handles file navigation ( StartupHub.ai ). Both live in Pierre's Apache-2.0 monorepo on GitHub, which shows about 6,400 stars. OpenAI says the projects will stay open source and that it will keep contributing to them ( StartupHub.ai ).

Then the commercial side. Code Storage describes itself as "a high-scale, distributed storage system built on Git" with "API ergonomics optimized for machines." It serves native Git clone, push and fetch, ships SDKs for TypeScript, Python and Go, sends webhooks, and syncs with GitHub, GitLab and Bitbucket. The customers on its homepage include Lovable, Bolt, Amp, Poke, Anything and Parahelp ( Code Storage ).

OpenAI plans to move Code Storage's features into Codex after the deal closes ( StartupHub.ai, ShiftDelete.Net ). Earlier the same week, OpenAI described Pando, its in-house source control system for engineers and agents working in parallel ( StartupHub.ai ).

Axios, which reported the deal first, framed it as OpenAI going after Microsoft's GitHub ( ShiftDelete.Net ). OpenAI's own post doesn't use that framing; it talks about adapting Git workflows for agents ( StartupHub.ai ).

Pierre brings two things to OpenAI: the diff and file-tree libraries already in Codex, and a Git host other AI products store code in.

According to OpenAI, Code Storage customers can keep using the service after the deal closes, while the team moves its features into Codex ( StartupHub.ai ). The announcement gives no closing date and no shutdown date ( ShiftDelete.Net ). It also says nothing about pricing after the move, which StartupHub.ai lists as an open question.

So nothing changes for you today. What you can do now is find out how much of your code would be easy to take somewhere else, and Code Storage's own docs answer most of that.

Code Storage can sync with GitHub in two modes. In GitHub App mode the sync is two-way, and the docs say "Code Storage treats GitHub as the source of truth for synced repositories." Public mode imports a public repo once and keeps nothing in sync after that ( Code Storage docs ).

Even in two-way mode, only refs/heads/* and refs/tags/* sync from GitHub, and ephemeral refs stay in Code Storage ( Code Storage docs ). The ephemeral namespace exists for "previews, CI artifacts, and experiments" ( Code Storage docs ). You reach it by adding +ephemeral before.git in the repo URL. The docs are direct about what that means: "A normal git clone does not show the ref," and "Code Storage never mirrors an ephemeral ref to an upstream host" ( Code Storage docs ).

Git LFS has a related catch. LFS files uploaded only to an ephemeral ref never reach GitHub, and promoting that ref through the API leaves pointers on GitHub with no bytes behind them ( Code Storage docs ).

If your product keeps agent previews or experiment branches in that namespace, those refs exist with one vendor. That was true before November 4. A pending acquisition with no timeline is a good reason to check it now.

A normal clone and a GitHub sync see the branches and tags. The ephemeral namespace stays below the waterline.

Start by listing what lives only in the ephemeral namespace. Code Storage remotes are HTTPS URLs authenticated with a JWT ( Code Storage docs ), so keep the full URL in an environment variable loaded from a secrets file, not typed into your shell history:

# URL ends in +ephemeral.git git ls-remote " $CODE_STORAGE_EPHEMERAL_URL " Enter fullscreen mode Exit fullscreen mode To copy those refs, try a mirror clone of the ephemeral URL into a second bare repo. The docs only show git push and git pull against that remote, so check that the clone actually holds the refs you listed:

git clone --mirror " $CODE_STORAGE_EPHEMERAL_URL " app-ephemeral.git git -C app-ephemeral.git show-ref Enter fullscreen mode Exit fullscreen mode For the normal namespace, the standard Git mirror recipe from GitHub's docs works against any Git remote. Clone once, point pushes at a backup you own, then fetch and push on a schedule:

git clone --mirror " $CODE_STORAGE_URL " app.git cd app.git git remote set-url --push origin " $BACKUP_URL " # repeat on a schedule git fetch -p origin git push --mirror Enter fullscreen mode Exit fullscreen mode The clone writes the URL, token included, into app.git/config, and Code Storage's Git tokens are JWTs, which can carry an expiry. Before each scheduled fetch, mint a fresh token and run git remote set-url origin "$CODE_STORAGE_URL", or hand the token to a credential helper.

A mirror push doesn't carry Git LFS objects. GitHub's docs add two commands for that:

git lfs fetch --all git lfs push --all " $BACKUP_URL " Enter fullscreen mode Exit fullscreen mode I haven't run any of these against Code Storage myself, the ephemeral clone included. They are plain Git, and they copy only the refs the remote shows you, which is why the ephemeral listing comes first.

Pierre also documents a full export into an S3 bucket you own: one bare repository per repo, "with all refs and history, plus every Git LFS object." You arrange it with Pierre's support, you stop reads and writes to your repos while it runs, and it excludes settings, access keys and webhooks ( Code Storage docs ). The export page doesn't say whether ephemeral refs are included. Ask before you rely on it.

Trees and Diffs are already how Codex shows you what an agent changed ( OpenAI, StartupHub.ai, Investing.com ). So the review step is where you spend your time, whoever hosts the repo.

I build Hermes IDE, so weigh this with that in mind. Hermes runs Codex, Claude Code and other agent CLIs each in its own terminal, and starts each task on its own hermes/<name> branch and worktree in your repo ( GitHub ). The Review Desk (⌘G, or Ctrl+Shift+G on Windows and Linux) shows an agent's changes by turn or by file, and Land commits, opens a pull request or merges locally, with undo. The branches stay in your repo, and you push them to whatever remote you already use. Hermes is free to download, source-available under BSL 1.1, on macOS, Windows and Linux, and not affiliated with OpenAI or Pierre.

The Review Desk (⌘G) in By turn view for a Claude Code task in orders-api: two turns, turn 1's diff of src/db.ts, the changed files, and a commit message drafted from the turns.

Three things decide whether this matters for your stack: the closing date, what Code Storage costs once it moves into Codex, and whether it stays a standalone API for other companies' products. StartupHub.ai also asks how Pando and Code Storage will fit together ( StartupHub.ai ). OpenAI hasn't answered any of them yet.

If your product stores user code with a Git host built for agents, do you keep a mirror somewhere else?

Code Storage sells itself on system-design promises: quorum writes, failover, a 99.99% SLA. Deciding where your source of truth lives, and what copy survives when a vendor changes hands, is the same building-block thinking the System Design Pocket Guide: Fundamentals walks through. It is one of 8 short books in the Pocket Guides for Developers series, covering agents, RAG, prompts, observability, events, databases and system design: see the whole series.

If you run more than one coding agent, Hermes IDE keeps each task on its own branch in your repo; it's built in the open on GitHub.