llms.txt
# Vexi
> Open-source AI coding agent for your terminal. Bring your own key, zero config, 100% local.
Vexi is a free, open-source CLI tool that brings AI coding assistance directly into your terminal. It works with any API key across 13 providers (Anthropic, OpenAI, Groq, Google Gemini, OpenRouter, Cerebras, DeepSeek, Zhipu AI, Qwen, Kimi, MiniMax, Mistral, Z.ai) or any OpenAI-compatible endpoint URL, runs entirely on your machine, and requires no account or server.
## Install
```bash
npm install -g vexi-cli
vexi
```
## Key features
- **BYOK (Bring Your Own Key)**: Auto-detects your provider from the key format. Supports 13 providers: Anthropic Claude, OpenAI, Groq, Google Gemini, OpenRouter, Cerebras, Zhipu AI (GLM), DeepSeek, Alibaba Qwen, Kimi (Moonshot), MiniMax, Mistral, and Z.ai.
- **URL-based setup**: `vexi setup` takes any OpenAI-compatible endpoint URL: (Ollama, LM Studio, custom proxies) and Vexi detects the provider, fetches available models automatically, and verifies the connection with a real request before saving the config.
- **Context Compression Engine**: Keeps a running summary of your decisions across sessions. Never re-explains your project.
- **Full project understanding**: Scans the entire codebase on startup (respects .gitignore, skips node_modules) and injects a compact map into every AI prompt.
- **Multi-language build support**: Runs pip install, gcc, javac, cargo, gradle, and more directly from chat. Works with Python, Java, C, C++, Rust, Go, and JavaScript.
- **Multilingual**: UI and code explanations in Arabic (ar), English (en), Spanish (es), Portuguese (pt), French (fr).
- **Session replay**: Every session recorded locally; export as interactive standalone HTML.
- **Custom skills**: Drop `.md` files in `.vexi/skills/` to teach Vexi your conventions. Share skills via GitHub URL, with a preview and confirmation before any remote content is trusted.
- **Visual dependency graph**: Interactive d3 force-directed graph exported as a single HTML file.
- **MCP support**: Works as both an MCP client (connects to external tools) and MCP server (exposes Vexi capabilities to Claude Desktop, Cursor, etc.).
- **Vexi Learn**: Mines past sessions for your corrections and builds a personal style skill file.
- **Git push from chat**: `/push` stages, drafts an AI commit message, and pushes without leaving the session.
- **Reliable file edits**: Structured `read_file`/`write_file`/`edit_file` tools replace fragile shell patching, giving exact substring edits with precise per-file undo snapshots. Uses native function-calling on providers that support it, falling back to a provider-agnostic text protocol otherwise.
- **Token & cost tracking**: Real per-session token usage with an estimated dollar cost. Check anytime with `/usage`, plus a summary at exit. Known free-tier models read as $0.
- **Undo / redo**: Every AI file edit is snapshotted before it happens; `vexi undo`, `vexi redo`, and `vexi history` roll back or replay changes without touching git.
- **Non-interactive print mode**: `vexi -p "<prompt>"` runs a single non-interactive turn for scripts and CI pipelines; add `--yes` to auto-run any shell commands it proposes.
- **Self-maintaining**: `vexi update` and `vexi uninstall` (with `--purge`), plus a daily background check that notifies you of new versions.
## Privacy
- Your API key is stored locally at `~/.vexi/config.json` (chmod 600). Never sent to Vexi servers.
- Your code is sent only to the AI provider whose key you supply.
- No telemetry, no analytics, no accounts.
## Links
- Website: https://vexi.pro
- Documentation (5 languages): https://vexi.pro/docs.html
- Blog: https://vexi.pro/blog/
- About: https://vexi.pro/about.html
- Contact: https://vexi.pro/contact.html
- Privacy & Terms: https://vexi.pro/privacy.html
- npm: https://www.npmjs.com/package/vexi-cli
- GitHub: https://github.com/Elomami1976/vexi
- Email: hello@vexi.pro
- License: MIT
## Blog
- [Vexi vs Claude Code, Cursor and OpenCode](https://vexi.pro/blog/vexi-vs-claude-code-cursor-opencode.html): An honest look at four AI coding tools, including the cases where Vexi is the wrong choice and one of the others is clearly better.
- [How to choose an AI coding agent, without the marketing](https://vexi.pro/blog/choosing-an-ai-coding-agent.html): Five questions that narrow the field faster than any feature table, and what each answer rules in or out.
- [Why Vexi exists: an AI coding agent that never leaves your machine](https://vexi.pro/blog/why-vexi-exists.html): Most AI coding tools ask you to trust a server with your source code. Vexi takes the opposite path: no account, no backend, no telemetry.
- [How Vexi detects 13 AI providers from your key alone](https://vexi.pro/blog/bring-your-own-key.html): No provider dropdown, no config file. The format of an API key is enough to identify who issued it. Here is how that works.
- [Running an AI coding agent for $0: the free-tier providers](https://vexi.pro/blog/free-ai-coding.html): Eight of the thirteen providers Vexi supports have a genuine free tier. Here is what each is actually good for, and where the limits bite.
- [How Vexi remembers your decisions across sessions](https://vexi.pro/blog/context-compression-engine.html): Context windows end. Projects do not. The Context Compression Engine keeps what matters and drops what does not.
- [From prompt to pushed commit without leaving the session](https://vexi.pro/blog/git-push-from-chat.html): The /push command stages your changes, writes a commit message from the real diff, and pushes, with a confirmation you actually control.
- [Knowing what a session actually cost, while it runs](https://vexi.pro/blog/token-and-cost-tracking.html): BYOK means the bill is yours. /usage shows real token counts and an estimated dollar cost per session. No dashboard, no lag.
- [Teaching Vexi your conventions with custom skills](https://vexi.pro/blog/custom-skills.html): Drop a Markdown file in .vexi/skills/ and Vexi follows your house rules. No fine-tuning, no prompt engineering in every message.
