Home/Terminal AI Coding Agents/Vexi
V

Vexi

Updated: Nov 16, 2026

Install an open-source AI coding agent with one npm command, keep your keys and code on your machine, and push commits without leaving the terminal

Image not available

About Vexi

Overview

Vexi is an open-source AI coding agent that runs in your terminal. You install it with a single npm command, point it at the API key you already own, and keep your code and keys on your own machine. The project ships under the MIT license and sits at version 0.9.2.

It scans the entire project instead of the open file, carries decisions across sessions, and handles staging, commit messages and git pushes from inside the chat. Vexi also explains code in Arabic, Spanish, Portuguese, French and English, and runs build commands for Python, Java, C, Rust and Go.

Key Benefits

  • One npm install covers Windows, macOS and Linux, with no config files to write.
  • You bring your own key: Vexi auto-detects the provider from the key format and stores it locally only.
  • The Context Compression Engine keeps a running summary of your decisions so you stop re-explaining the project.
  • Every AI file edit is snapshotted first, so /undo, /redo and /history make experimenting safe.
  • Git operations happen inside the chat, with credential pre-flight checks before each push.
  • /usage reports real token counts and an estimated dollar cost for the session.

How It Works

Run npm install -g vexi-cli, then open a session inside a project directory. Vexi scans the codebase, loads stored memory and waits for a prompt. You describe the change, it locates the file and proposes an exact edit with a snapshot saved beforehand. Type /push to stage the change, review the AI-drafted commit message and confirm the push to origin.

Use Cases

  • Terminal-first developers fix bugs, patch files and run builds from a single prompt loop without leaving the shell.
  • Engineers dropped into unfamiliar repositories rely on whole-project scanning and stored decisions to get oriented faster.
  • Polyglot teams run pip install, cargo, javac or gcc straight from chat across Python, Rust, Java, C and Go work.
  • CI and automation engineers call vexi -p with a prompt for a single non-interactive reply, adding --yes to auto-run proposed commands.
  • Developers working in Arabic, Spanish, Portuguese or French read code explanations in their own language.
  • Privacy-conscious teams keep everything on local machines, since no server is involved and keys never leave the computer.

Why Choose This Product

Vexi competes on constraints rather than breadth: MIT-licensed, no server required, and compatible with keys from 13 providers, including free-tier models at Groq, Gemini, DeepSeek, Cerebras, Qwen, Kimi, MiniMax and Zhipu. Its own comparison table sets it against OpenCode, Claude Code and Cursor on install simplicity, provider freedom, context memory and local execution.

Vexi Pros & Cons

Strengths
  • Open source under the MIT license at no cost
  • One-command npm install on Windows, Mac and Linux
  • BYOK across 13 providers, including free-tier models
  • Runs 100% locally with no server involved
  • Builds and runs code for Python, Java, C, Rust and Go

Key Features

๐Ÿ”‘

Provider Auto-Detection

Vexi reads the key format to identify the provider automatically, or accepts any OpenAI-compatible endpoint URL via vexi setup, and stores the config locally only.

๐Ÿง 

Context Memory

A Context Compression Engine keeps a running summary of project decisions and injects it into later sessions so you stop re-explaining the codebase.

๐Ÿ“‚

Whole-Project Scanning

Vexi scans the entire codebase rather than the open file, respects .gitignore and skips node_modules.

โ†ฉ๏ธ

Undo And Redo

Each AI file edit is snapshotted first, so /undo reverts it, /redo reapplies it and /history browses the full change log.

๐Ÿš€

Git Push In Chat

The /push command stages files, drafts an AI commit message and pushes to origin, with credential pre-flight checks before every push.

๐ŸŒ

Multilingual Explanations

Vexi explains any code in Arabic, Spanish, Portuguese, French or English.

๐Ÿ› ๏ธ

Multi-Language Builds

Vexi auto-executes build commands such as pip install, gcc, javac and cargo for Python, Java, C, Rust, Go and other languages directly from chat.

๐Ÿ“Š

Token And Cost Tracking

The /usage command shows real input and output token counts plus an estimated dollar cost per session, and exit prints a summary.

๐ŸŽฏ

Personal Skill File

Vexi mines past sessions for your corrections and builds a personal skill file that it injects into every future session.

๐Ÿค–

Scriptable Single Replies

vexi -p with a prompt returns one non-interactive reply for scripts and CI pipelines, and --yes auto-runs any commands it proposes.

Frequently asked questions about Vexi

Is Vexi free to use?

Yes. Vexi is free and open source under the MIT license, and the page describes it as yours to keep. You install it with npm install -g vexi-cli and supply your own AI provider key, so model usage is billed by whichever provider you choose. Several listed providers offer free-tier models that Vexi reads as $0.

Does my code or API key leave my computer?

No. Vexi runs 100 percent locally with no server involved. Your API keys are stored locally only, and your code never leaves your machine.

Which AI providers does Vexi work with?

Vexi supports 13 providers: Groq, Google Gemini, DeepSeek, Cerebras, GLM Zhipu AI, Qwen Alibaba, Kimi Moonshot and MiniMax, plus BYOK options Anthropic Claude, OpenAI, OpenRouter, Mistral and Z.ai. It auto-detects the provider from the key format, or you can paste any OpenAI-compatible endpoint URL.

What operating systems does Vexi run on?

Vexi works on Windows, Mac and Linux. It installs globally as vexi-cli with npm and requires no config files.

How do I undo an AI edit?

Vexi snapshots every AI file edit before it happens. Type /undo to roll back the change, /redo to reapply it, and /history to browse the full change log.

Can I use Vexi in CI pipelines or scripts?

Yes. Running vexi -p with a prompt returns a single non-interactive reply suited to scripts and CI pipelines. Adding --yes lets it auto-run any commands it proposes.