· via dev.to (home feed)
Gitpod, now Ona, halts new signups and repositions cloud dev environments around Codex
Gitpod, now operating as Ona, has closed new customer signups and points newcomers to Codex, one sign of a broader shift of hosted dev environments toward AI agent compute.

Gitpod stops onboarding new customers
Gitpod has stopped accepting new customers, and its documentation now opens by directing newcomers to Codex rather than touring the product. According to a post on dev.to by Marcus Beauregard, the company — which for many developers defined what a cloud development environment is — now operates under the name Ona, and what it sells is hosting for background agents rather than environments for people.
One vendor changing course would be a press release, the author concedes. But after reading the documentation of 25 projects across the space, from bubblewrap to Cloudflare Sandboxes, he concludes the shift spans the entire layer: the buyer has changed, and most maps of the dev environment market, including his own, are stale.
Four vendors, one direction
The post lines up four moves as evidence. Ona, formerly Gitpod, has closed signup and points its docs at Codex. Daytona has moved from dev environments to sandboxes for code written by AI, with its own kernel, OCI compatibility and a claimed start time under 90 milliseconds. Dagger presents itself as a third generation of CI, succeeding Jenkins and GitHub Actions, and built to keep pace with coding agents. Earthly has recast its build tool as Earthly Lunar, framed as guardrails for AI-era engineering. The author reads three of these as the same strategic move and the fourth as that move with a bleaker outcome, and he flags that every number cited is a vendor claim he did not benchmark.
A four-layer map of the space
To explain why arguments in this area go in circles, the post sorts it into four layers:
- Kernel isolation (namespaces, seccomp, gVisor, Firecracker): how you stop a process.
- Runtimes and images (Docker, Podman, Kata, OCI): how you ship a filesystem and entrypoint.
- Declaration (Nix, devcontainers, mise, uv): how you state what a project needs.
- Hosted environments (Codespaces, Coder, Daytona, E2B): who runs it for you, per session.
The layers do not stand in for one another. A hosted tool at layer four solves nothing at layer one; a declaration tool at layer three assumes layers one and two are already handled. A sandbox that boots in under 90 milliseconds says nothing about whether your Postgres is reachable, which is why tools from different layers keep getting debated as though they competed.
Solved: versions and hermetic builds
Two problems are largely settled, per the post. Tool version management has consolidated: mise folds asdf, direnv and a slice of make into one file; uv replaced pyenv, poetry and tox for the author in an afternoon; Pixi covers the conda side where native libraries matter; Hermit's committed shims offer onboarding that works straight after cloning. Hermetic builds are also served: Nix remains the only genuinely reproducible option, with flake.lock pinning the input graph being the crucial part, and devenv, Devbox and Flox acting as friendlier interfaces over the same underlying store.
Unsolved: services and trust in downloads
Two problems are not. Services and system libraries are still dodged: most declaration tools either punt to Docker or grow a process manager, with devenv cited as the exception that actually declares Postgres and Redis with lifecycle management while others declare versions and assume libpq is already present. Trust in downloads is also unresolved, because declarations that can point at arbitrary URLs can also be rewritten by a malicious pull request, which is where most supply-chain risk gets glossed over.
The environment as an input
The post's author builds a project that runs repositories on a host it does not own, so he treats the environment as a named input rather than scenery. An env-check command verifies that a repository's declared environment is present and intact; env-init builds it, printing a plan unless given an explicit apply flag. Environments are keyed by digest and built in a temporary sibling directory that is renamed only on success, so a failed init leaves nothing a later check would accept. What can be downloaded comes from the runner's list — one pinned URL and checksum per toolchain — because a declaration picks a version, never a source. Child processes receive an allowlisted environment plus a scratch home and temp directory inside the run, since a child that inherits the parent's environment can read every credential on the machine.
Why it matters
The open question, per the post, is who owns the environment when the person who wrote the code is not the person who ran it. A host executing someone else's code has to name the environment the way it names a branch or a base commit, and reject unsuitable input up front instead of finding a half-installed toolchain mid-run and reporting it as a test failure. The layer-four vendors already grasp this, the author argues — it is why they have turned away from human customers. For teams, the practical guidance is to judge environment tooling by layer rather than as interchangeable options, to pin toolchain sources at the runner instead of trusting repository declarations, and to expect hosted dev environments to keep consolidating around compute for agents.
- #gitpod
- #cloud-development
- #ai-agents
- #developer-tools