We implemented Kimi K3's native harness in Rust.
Now, @OpenInterpreter is the #1 trending Rust repo in the world.
Apache licensed, ACP + Codex SDK compatible — a CLI for the next generation of open models.
We implemented Kimi K3's native harness in Rust.
Now, @OpenInterpreter is the #1 trending Rust repo in the world.
Apache licensed, ACP + Codex SDK compatible — a CLI for the next generation of open models.
@_xonoxc Totally flexible, you can use any harness for any model! In the interactive TUI you run /harness, very similar to running /model. You can also specify it for headless runs.
When we say best: Each model just has a default harness, usually the one built by the provider.
Open Interpreter provides everything the Codex CLI provides (it's a fork of Codex, broadened to work well with open models).
That includes:
- A mature terminal interface
- Great sandboxing on Windows/Mac/Linux
- An app-server that supports the Codex SDK
- The usual stuff (skills, MCP, subagents, etc.)
This particular emulation of Kimi Code lets Kimi think its running inside its native harness, which it seems more comfortable in.
More details on why we're doing harness emulation:
https://t.co/No4BBzz3KK
1. Yes, you can sign in with oauth subscriptions like Kimi Code and ChatGPT. Our goal is to support any token source, including API keys or oauth. (As long as TOS allows it — Anthropic is the exception here.)
2. It does have the Codex app server! Open Interpreter is a fork of Codex, so it has a server mode, Codex SDK compatibility, etc. Docs: https://t.co/FmQ77oxlTq
3. Same as Codex, it works well headlessly with a command like `interpreter exec -c model='"kimi-k3"' \ "your prompt here"`. Docs: https://t.co/RX0ZenalbT
Open Interpreter is a drop-in replacement for Codex.
It's also ACP compatible, with Claude Agent SDK compatibility on the way.
However you build with agents, we want to support your switch to open models.
We're on a mission to implement the best harness for each model in one coding agent.
For each model, the most performant harness is used under the hood.
Follow our blog for details:
https://t.co/4WCnCf8FaS
Why reimplement a coding harness?
Because harness choice dramatically affects a model's performance. Some models are even trained inside a specific harness.
Now that the provider ecosystem is diversifying, we needed an agent core that could adapt its harness.
Pretty cool to see a popular fork of Codex OSS repo with 60k stars out there with my team being top contributors (due to preserving history with the fork)
Keep building
https://t.co/ovpDi3UtiT
Today I'm launching Interpreter.
It's a desktop agent that can fill PDFs, edit your Excel and Word docs, and learn new skills.
Runs offline, works with any model, and it's free.
Today we're discontinuing the 01 light, refunding everyone, and launching a free 01 app.
We're also open-sourcing all our manufacturing materials + a major 01.1 update.
Why? Focus. This software will change the world. That deserves our full attention.
Also phones
🧵/8