Version 0.31.0 puts many model providers behind one terminal
OpenClaude v0.31.0 keeps the familiar coding-agent loop: ask for a change, let the model inspect files, approve shell or editing tools, and review the result. The same terminal can use OpenAI-compatible APIs, Gemini, GitHub Models, Codex authentication, Ollama, LM Studio, and several named gateways. MCP, subagents, slash commands, background sessions, and a VS Code extension sit around that loop. Provider switching is the product, rather than one more chat window for a single model company.
A provider swap does not promise identical behavior. The README says Anthropic-specific features may be absent elsewhere and smaller local models can struggle with long tool sequences. Its Ollama route requests a 32,768-token context window to avoid the smaller default used by Ollama's compatibility layer. That is a sensible fix for one failure mode, yet the model still determines how well tools are chosen and whether a multi-step edit stays on course. You need to test the provider and model combination you intend to keep.
What happened when we ran it
Our sandbox installed commit 53c7ed9 in 35 seconds, adding 596 packages and using 446 MB on disk. The fresh checkout contained 3,499 files, about 927,757 lines of source, and occupied 34.9 MB. We used an unprivileged Debian container with 3 CPUs, 8 GB of RAM, Node and Bun, and no secrets. Installation succeeded without a provider account because this stage only prepared the repository.
The build passed in 24 seconds, and the supplied tests passed in 196 seconds. The repository also had 3 CI workflow files, a Dockerfile, and a tests directory. Those are useful engineering signals, especially in a codebase near 1 million source lines. Our run did not connect a model, judge generated patches, compare provider quality, or expose the gRPC service. The result proves that commit 53c7ed9 installed, built, and passed its test command in our container. It does not measure coding accuracy.
Node.js 22 gets you started, while each provider adds its own setup
The published package requires Node.js 22 or newer. Building from source adds Bun 1.3.13 or newer, and the README has a separate check for ripgrep if the installed command cannot find it. OpenClaude does not read a project .env by default. The guided /provider command stores a profile, while shell variables or --provider-env-file cover manual setups. This is clear documentation, although it means copying an existing Claude Code setup will not make the tool ready.
OpenClaude keeps its files under .openclaude and explicitly avoids Claude Code's .claude paths. The migration guide says to copy only settings, commands, agents, skills, or scheduled tasks that you created, then configure authentication again. Local Ollama and LM Studio routes avoid a hosted API key but require a running server and a downloaded model. Hosted routes bring their own key, account, pricing, and output limits. One interface cuts switching work, but provider operations remain.
The LICENSE leaves the core distribution rights unresolved
The v0.31.0 package declares SEE LICENSE FILE, and that file is unusually direct. It says OpenClaude contains code derived from Anthropic's proprietary Claude Code CLI. Contributor modifications are offered under MIT where legally permissible, while the underlying derived code remains under Anthropic's copyright. The notice also says Anthropic has not authorized distribution and tells users and contributors to evaluate their legal position. GitHub consequently reports the license as NOASSERTION, rather than calling the whole repository MIT.
That notice matters more than the 35-second install. A developer can inspect the code and decide whether a personal experiment fits their tolerance. A company embedding it in an internal platform, shipping it to customers, or building policy around it has a different exposure. We cannot settle that legal question for you, and the project does not pretend it has. The practical recommendation is plain: treat the LICENSE as an adoption gate, not a footer to read after deployment.
October 2026 activity is high, but support follows the newest release
GitHub showed 33,677 stars, 40 open issues, and 28 open pull requests on October 8, 2026. The last push was October 7, after the commit our lab ran, and changed file-read deduplication. Release v0.31.0 shipped on September 22 with provider additions and fixes across permissions, background sessions, context handling, MCP, images, and OAuth state validation. The combination of a recent push and current issue work points to active maintenance.
Issue 2255 reports that the Arch community package could not start because a launcher helper was missing, and pull request 2256 proposed a fallback on October 5. The security policy supports only the newest npm release and current main branch, not older versions or modified builds. If you adopt OpenClaude, stay close to upstream, pin what you deploy, and recheck packaging on your operating system. Active does not mean every distribution path is equally settled.
A 596-package clean run cannot answer the license question
Our 596-package install and successful 196-second test run make OpenClaude easy to trial on technical grounds. Its provider range is useful when you want one agent interface for a local model today and a hosted service tomorrow. The docs spell out the limits: provider behavior varies, configuration lives separately, plaintext routing keys need care, and the headless server needs protection.
The LICENSE is still the deciding document. Choose OpenClaude for a controlled evaluation only when its Claude Code-style workflow is the specific thing you need and you accept the work of testing each provider. For organizational use or redistribution, get a legal answer before investing in integrations. OpenCode, Aider, or Codex give you a cleaner place to start when that answer is uncertain.

