Four interfaces, with a model service you choose
Reasonix is worth considering if you want terminal, desktop, browser and editor access to the same coding engine. The current README describes an agent that reads code, edits files and runs commands in a selected project. A DeepSeek preset is included, and an OpenAI-compatible service can be configured.
The local engine and the model are separate. The README explicitly says the configured model service receives what the agent reads. Provider choice alone establishes neither lower costs nor better coding results; our run measured neither.
The README labels Reasonix 1.x on main-v2 maintenance/stable and Reasonix 2.x on studio active development. At our measured commit, the README described 2.x releases as prereleases. The current README instead says release candidates such as -rc.N are marked as prereleases. Studio v2.23.0 was a prerelease; that historical label should not be applied to every current 2.x release.
What our September 28 run actually establishes
Our recorded run began on September 28, 2026, at commit ae27da71e2a85decdb92bc9959f8ad05ab076b74 (ae27da7). The recorded sandbox configuration was 3 CPUs, 8 GB of memory and image golang:1.24-bookworm. An image label does not establish the toolchain actually used.
The install-labelled step exited 0 after 51.519 seconds, rounded to 52 seconds in the summary. Its retained output is go: downloading go1.26.8 (linux/amd64). The dependency summary records 115 entries, but this does not prove 115 application packages were installed or that the Reasonix CLI was installed.
The build-labelled step exited 0 after 146.149 seconds, rounded to 146 seconds. Its retained output is vet-exit:0. This establishes successful completion of that recorded step. Without its full command or an executable check, it does not establish that a CLI artifact was produced or launched.
The Go test command reached the 900-second limit and exited 124. Its parser recorded 76 successful package results, 4 failing package results and 3 packages with no tests. The stored summary also contains a total of 80. These figures are parser summaries, not individual test counts, and have not been reconciled against a complete log to establish exhaustive unique coverage. They cannot support a claim that Reasonix passed 76 of 80 tests.
The retained tail shows failures in reasonix/internal/session/control and reasonix/internal/state/migration. Three named migration tests failed, with output referring to lock config edits: resolve OS user: user: unknown userid 1000. That gives a concrete environment-related error to investigate. It does not explain every failing package or the overall timeout, and the failed command does not establish that the product is broken.
The checkout scan recorded 5,684 files, approximately 803,650 source lines, 46.3 MB and 22 CI workflows. It found no Dockerfile or directory named tests; the visible Go test output demonstrates why that directory scan must not be read as an absence of tests. No new execution, installed-binary check, model-backed coding workload, cost comparison or security audit accompanies this correction.
Setup depends on the client
The current README requires Go 1.25+ for CLI source builds. It documents a pinned toolchain and GOTOOLCHAIN=auto; the recorded Go 1.26.8 download is consistent with a toolchain download, not evidence of Go 1.24 compatibility. Studio adds Node 24+, pnpm 10 and platform webview dependencies.
Packaged Studio users connect a provider and model inside the app. Terminal users run reasonix setup, then reasonix. The documented 1.x CLI is available through npm or Homebrew on macOS. The VS Code extension starts a local reasonix acp backend and does not bundle the CLI.
The configuration guide says provider entries name secrets through api_key_env, while values live in Reasonix's global .env. Project .env files can supply workspace plugin values but are not provider-key runtime fallbacks. Provision credentials for the host that will run the engine, then verify a provider connection before expecting coding work to succeed.
Permissions and rewind need separate checks
The README describes approval levels and plan mode. The guide supplies ask, allow and deny permission configuration and workspace controls. These are documented capabilities; our run did not independently validate their enforcement.
The September 29, 2026 CLI v1.39.5 release notes describe stricter handling of project configuration and approval for project-declared programs. That release followed our measured commit, so our September 28 run is not validation of those changes.
File rewind has a narrower scope than restoring an entire task. The current README says it covers files changed by edit tools, not shell commands. Keep Git or another recovery method for shell side effects, database writes and deployments. Check recovery on disposable files before relying on it during a larger task.
Windows startup evidence remains unresolved
Issue 11434 began with a Windows 11 report about Studio 2.23.0 failing on first launch with the kernel sent no handshake, then opening on another attempt. Later discussion matters: a startup wait and retry change was announced for 2.24.0, followed by another report involving 2.33.0.
On October 18, maintainers said bounded shell discovery first shipped in Studio 2.37.0, but described it as a likely cause rather than a confirmed explanation. The owner's Windows test did not reproduce the original cold-start failure. Those are upstream reports and investigations, not our measurements. They establish neither that every current Windows build fails nor that the original case is resolved. Test the release you plan to deploy after a restart or update.
How to read the scores
These are editorial evidence ratings on a 1 to 5 scale: 1 means a documented obstacle, 2 means substantial uncertainty, 3 means a usable documented route with material limits, 4 means detailed support for that dimension, and 5 requires broad corroboration. Unknown adoption, compatibility and runtime reliability do not earn high scores.
Setup receives 3 because packaged and source routes are documented, while our retained run lacks a runnable-binary check. Documentation receives 4 for concrete client, toolchain and credential instructions, without claiming every documented workflow was tested. Community receives 3 for visible reporting and maintainer follow-up, not for stars or assumed adoption. Maturity receives 3 because a stable maintenance line exists alongside active 2.x development, with current reliability still requiring release-specific checks.
Decide with the release you will use
Codex CLI is a relevant alternative when you want OpenAI's local coding agent and its documented ChatGPT sign-in route. Claude Code is relevant for its documented terminal and IDE workflow, including code explanations, routine tasks and Git work. Their supplied READMEs support those roles, but provide no comparative performance or reliability result against Reasonix.
For Reasonix, choose a release and client, configure the provider, confirm the executable launches, and try a bounded coding task with explicit checks. Verify edit-tool rewind separately from shell recovery. Our historical source run identifies questions to investigate; it does not settle whether today's release fits your work.