An open-source agent for the models you run yourself
English · 简体中文 · 繁體中文 · 日本語 · 한국어 · Español · Português · Français · Deutsch · Русский
Download · Documentation · Getting started · Providers · Limits
OpenCLI is an agent that runs on your own computer and works with any OpenAI-compatible model — one on the machine in front of you, one on your own server, or a hosted provider when you want one. It reads your files, runs your commands and edits your work, in a terminal and in a desktop app that share the same conversations.
Nothing is built in. No models, no accounts, no keys, and no opinion about where your inference comes from. Point it at an endpoint and its models appear on their own; what you type goes there and nowhere else.
| Platform | |
|---|---|
| macOS · Apple Silicon | OpenCLI-macos-aarch64.dmg |
| macOS · Intel | OpenCLI-macos-x86_64.dmg |
| Windows | OpenCLI-windows-x86_64-setup.exe |
| Linux | OpenCLI-linux-x86_64.AppImage |
The app updates itself afterwards. These builds carry no Apple or Microsoft certificate, so the first launch needs one extra step — what to do.
Command line, macOS and Linux:
curl -fsSL https://opencli.ai/install.sh | shThen point it at a model.
-
🗂 Works on files, not just code. It reads a folder of invoices the same way it reads a repository — and most of the work people actually have is the first kind.
-
🏢 Organised like an office. A department is a directory with standing instructions. A bot is a chat inside one, with a job. Nine departments ship with sample data, so each can be tried before it is set up.
-
⏰ Duties that come round on their own — and stop to ask you when they reach something they should not decide alone. That threshold is written in words, in advance: any single line over 10,000.
-
🤝 Bots hand work to each other, with what they did and the files they produced. Capped at eight hops and three appearances each, so a cascade cannot run away.
-
🧰 Six workflows ship as skills and work in any directory: reviewing rows against their rules, reporting what changed over a week, triaging an inbox, comparing two drafts, turning a transcript into decisions and owners, and reading records for what is worth checking.
-
🔌 Connectors are MCP servers — GitHub, Postgres, Slack, Notion, a browser. Keys are kept away from the configuration file, which gets shared.
-
🚦 A sandbox with a visible edge. A run may write inside its own directory and nowhere else, and a run pointed somewhere unfamiliar is held until you allow that place by name.
-
🌍 Ten languages, and any other is one JSON file you can add without rebuilding anything.
-
💻 Terminal, desktop and a local web UI over the same gateway, so a conversation started in one continues in another.
Not thirty integrations — one protocol. Anything answering /v1/chat/completions and /v1/models the way OpenAI does can be pointed at, and its models appear on their own. Every endpoint below was asked for its model list before being listed here.
On your machine — Ollama · LM Studio · llama.cpp · LocalAI · Jan · Xinference · KoboldCpp · llamafile
On your server — vLLM · SGLang · Text Generation Inference · NVIDIA NIM · LiteLLM
Hosting open models — Fireworks · Together · Groq · DeepInfra · Baseten · Novita · Hyperbolic · Nebius · Cerebras · SambaNova · Featherless · Chutes · NVIDIA · Hugging Face · OpenRouter · Vercel AI Gateway · Perplexity
In China — 硅基流动 · 无问芯穹 · PPIO · 七牛云 · 魔搭 · 阿里云百炼 · 火山方舟 / 豆包 · 百度千帆 · 讯飞星火 · 小米 MiMo · DeepSeek · 智谱 · 月之暗面 · MiniMax · 阶跃星辰 · 腾讯混元 · 百川
Closed models, when you want one — OpenAI · Anthropic · Mistral · xAI
Several give a free allowance, which is the cheapest way to try a model larger than your machine can hold. Endpoints, and which providers carry which model: Providers.
Reachable is not the same as good at this. Tool calling separates a model that can do this work from one that can only talk about it, and the model card never says. Check yours — one fixed task, five minutes. The table has one row because one model has been run, and adding a row is the most useful contribution you can make.
- General open models reach for
grepwhen they have a file tool. A gap in the models; no interface closes it. - Token counts are estimated, never counted. The tokenizer belongs to the model, and this runs models it has never seen.
- Scheduled work runs only while OpenCLI is open. A local agent, not a server.
- A background run may write anything inside the directory it runs in.
- The desktop builds are not signed. Your operating system is right to warn you.
- Right-to-left languages are not laid out yet.
Limits says why for each.
A fork of OpenAI's Codex CLI, Copyright 2025 OpenAI, Apache-2.0. Not affiliated with OpenAI.
Codex is built for OpenAI's models and this is built for everybody else's, which changes more than a name. Departments, bots, duties, background runs and the workflows are ours, as is a good deal of work on what breaks when the model is small — a compaction loop that summarised once a turn for eleven turns and threw away the thread each time, and a token estimate that read a page of Chinese as three quarters of its real cost.