This is a Codex fork for switching model harnesses
Open Interpreter has changed shape. The current repository is a Rust coding agent based on OpenAI's Codex, aimed at getting useful tool behavior from lower-cost and local models. The original Python project now lives in a community-maintained fork. Old tutorials and configuration examples may describe a different codebase.
The main control is /harness, which lists 9 documented choices including native, Claude Code, Kimi Code, Qwen Code, DeepSeek TUI, SWE-agent, and minimal modes. A harness shapes the instructions and tools around a model. Switching one can help compare how a provider behaves, but it also adds another variable when a command loop, context limit, or tool call goes wrong.
ACP and Codex compatibility reduce client lock-in
Open Interpreter can run as an Agent Client Protocol agent inside compatible editors. It also speaks the Codex exec protocol, and the README shows a Codex SDK client using codexPathOverride to launch interpreter. Shared AGENTS.md files, .agents/skills directories, and MCP integrations can be read without conversion into a private format.
Product-specific state still exists. Configuration, credentials, sessions, logs, caches, and daemon data remain under ~/.openinterpreter or an operating system credential store. The portability guide treats shared files and protocols as the reusable surface, while runtime history remains tied to the application. That is a reasonable boundary, but a team migrating agents should inventory both sides rather than assuming one directory contains everything.
What happened when we ran it
Our sandbox checked commit 5b07159 with 3 CPUs and 8 GB of RAM. The pnpm install succeeded in 40 seconds, added 1 package, and occupied 170 MB. The repository had 6,475 files, about 1,505,323 lines of source, and a 69.6 MB checkout. Its root package is a private maintenance workspace with Prettier as its development dependency.
There was no root build script or target, so our harness skipped build. There was also no test script or target, so tests were skipped. This is not a passing build or passing suite. It means the automated Node entry point could install its single package but had no command by which to judge the Rust application. Our scan found 25 CI workflow files, no Dockerfile, no tests directory, and monorepo workspaces.
Those facts make the published binary the faster evaluation route, while source buyers need to inspect the Rust workspace and CI commands themselves. The 40-second pnpm result verifies repo-wide maintenance tooling only. It does not exercise a provider login, a model response, command approval, MCP connection, ACP session, browser action, or Codex SDK request.
Provider and harness pairing can decide whether tools run
Open Interpreter can switch providers and models from its terminal interface, but compatibility has layers: provider, model, wire API, harness, and tool schema all have to agree. Hosted routes require the relevant API credentials. Local routes avoid a hosted key but still depend on a running model server and a model capable of following the selected harness.
Issue 1883 gives a concrete failure on version 0.0.40. The reporter could connect to Moonshot K3, yet prompts requiring shell or browser tools stopped with a compatibility-transport error before any tool ran. Issue 1889 describes an Ollama Qwen 2.5 Coder 7B setup repeatedly selecting an explorer target instead of filesystem or terminal tools. Each report concerns one configuration, but both make a staged provider test essential.
Start that test with read-only repository questions, then a harmless command, a controlled file edit, and a task that needs more context. Record the selected model, harness, and wire API with each result. A label such as OpenAI-compatible is too broad to prove that tool calls, streaming responses, context accounting, and error recovery agree with this client.
Context settings can regress on a specific provider path
Issue 1877 reports that a DeepSeek configuration on version 0.0.39 reset a user-specified context window after the first prompt and began compacting frequently. The report names Windows 11, the native harness, and the Responses configuration. It does not establish behavior for every provider or for release 0.0.40, but it identifies a useful acceptance check for long coding sessions.
Run a conversation past the point where compaction starts, inspect the displayed limit after the first response, and confirm that instructions survive. This matters more than a one-message demo because an agent can answer an opening prompt correctly while losing earlier constraints later. Our lab ran no model session, so it provides no evidence about output quality or context retention.
Command execution requires a local permission policy
The README says commands run inside native sandboxing on macOS, Linux, and Windows, and links to sandbox and approval documentation. That is better than treating shell access as an invisible side effect. Still, the useful safety boundary is the policy you configure on the platform you deploy, including network access, writable paths, MCP tools, and when a human must approve an action.
Computer use widens that boundary. The bundled QA skill can drive a browser through agent-browser or native applications through trycua. A team should keep initial trials inside disposable accounts and checkouts, especially when testing an unfamiliar 7B local model or a newly selected harness. Model price is not a substitute for reliable tool selection.
August 2026 activity accompanies a very new release line
GitHub recorded 68,206 stars, 9 combined open issues and pull requests, and a last push on August 20, 2026. Release 0.0.40 was published that day. Open reports from August 18 through 29 cover DeepSeek context handling, Kimi transport behavior, custom-provider documentation, and Ollama tool choice, so issue activity is current alongside the code.
Open Interpreter earns a trial when model and harness comparison is itself useful. Its shared files, MCP, ACP, and Codex protocol can reduce migration work between clients. The trade is a larger compatibility matrix than a single-provider agent has. Test the exact provider, harness, context length, permission policy, and tool set you intend to use before allowing it to work unattended.