Open Interpreter

Open Interpreter is now a Rust rewrite built on Codex, with native sandboxing and a provider list aimed at cheap and open models, Ollama included, and the daily trait is that a local model in a sandboxed shell costs nothing per token and asks before it runs anything. For scripting chores on your own machine that is a good afternoon.

What you will not love is the churn: the original Python project continues as a community fork, so the tutorials you find may describe a different program. Pick it for open models in a sandboxed shell. Pick Claude Code for no assembly and a model that needs fewer retries.

The worst thing is lineage. The current Open Interpreter is a Rust rewrite based on Codex, and the original Python project continues as a community-maintained fork, so two codebases share one name and an issue filed against one may describe the other. An MCP client exists, git operations do not, and no benchmark.

The consequence is that the roadmap belongs partly to Codex upstream, and a breaking change there arrives here on someone else's schedule. A clean split of the names would change this verdict. What it does right is the sandbox: commands run inside native sandboxing, with approval modes that control when it asks.

Three documented layers. First, execution: commands run inside native OS sandboxing on macOS, Linux and Windows, with approval modes governing consent, so verification is the user's approval and the sandbox's boundary. Second, the harness: the Rust rewrite is based on Codex and supports switchable harnesses, so the loop is a pluggable component rather than the product. Third, the model layer: Kimi, DeepSeek and Claude, plus Ollama locally.

No benchmark is published, so targeting open models is a design statement, not a result. The observation: a tool whose loop is swappable has declined to claim the loop is what makes it good.

Open Interpreter has the adoption signal every founder wants, 68,200 GitHub stars, and none of the monetization: BYOK only, pricing power zero because there is no price, and no pricing page to raise. The pivot already happened once, a Rust rewrite based on Codex, which is a team betting on someone else's harness to stay relevant.

Distribution without a business model is a feature waiting for a buyer who has one. Likely acquirer is an open-model provider, Moonshot or DeepSeek shaped, that wants a terminal in front of its weights. Position: use it freely, keep your prompts portable.

The demo is a shell agent that asks before it runs things, and approval modes are the control I want in a laptop policy, because they are the difference between a tool and a liability on sixty machines. Procurement has less to hold: no vendor, no seat price, no support agreement, no SSO, and no retention statement because nothing is retained by anyone but the model provider.

Fit with our pipeline is the strong part: an MCP client and a headless exec mode. Onboarding is a curl install. Not yet, though the sandbox and approval design is what I would ask the tools we do buy to copy.

This one respects me. Apache-2.0, a Rust binary, local models via Ollama or LM Studio autodetected, and a native sandbox with approval modes I choose per session. Switchable harnesses means the agent loop is pluggable, and I can mount it as an MCP server or an ACP agent, so my editor drives it and my other agents call it.

The grudge: [mcp_servers] is TOML I hand-write, and the name now covers two codebases, so a search lands on the wrong one. A fork survives the vendor, and the Python fork already proves it. Respect, with one open issue.