Wizard

Wizard is a self-extending terminal AI agent that ships as a single Rust binary. The whole pitch fits in a line, one command installs the agent, a model, and its tooling, and it works out of the box. From there it runs as a TUI in your terminal and talks to whatever model you point it at, whether that's an OpenAI-compatible endpoint, Anthropic, xAI, Groq, or a local model through llama.cpp. You can switch the live agent between providers on the fly with a command, so you're never locked to one vendor. The install is a single curl script, and the source lives on GitHub under a permissive MIT license.

The word that defines Wizard is self-extending. Through its evolve command, the agent can add new skills, MCP servers, scripted tools, and subagents to itself, and it does this by writing plain files that go live after a reload. Crucially, those changes are gated. A new capability only sticks if the project still compiles and passes smoke tests, and if something breaks the agent can roll back and log what happened. That turns the scary idea of an AI editing its own tooling into something with guardrails, because a bad change fails the build instead of silently corrupting your setup. It means the agent can grow to fit a task rather than being frozen at whatever its author shipped.

Beyond extension, Wizard has a few tricks that set it apart from a plain chat-in-the-terminal tool. Its fusion command runs a panel of your configured providers as a kind of debate and synthesizes a single tool-capable answer, which tends to beat any one model on its own. It supports the Model Context Protocol at runtime, so both stdio and HTTP MCP servers merge into its tool registry without a rebuild. There's a built-in benchmarking harness through the bench command that records and replays tasks, and a messaging gateway that can drive the agent from a Telegram bot. Project-level configuration and custom commands round out the day-to-day ergonomics. Local inference is a first-class option rather than an afterthought. Wizard can run models through llama.cpp and sizes GGUF files automatically, and it ships alternative install paths for Nix, minimal setups, and bring-your-own-model configurations. That flexibility is the point. You can start with a hosted frontier model when you want the best quality, then fall back to something local when you'd rather not send data out or pay per token, all without switching tools or relearning the workflow.

Wizard runs in a few distinct modes depending on how much autonomy you want to hand it. Genie is the interactive TUI for hands-on work where you stay in the loop. Sovereign is a headless, self-directing mode for when you want it to just go. Continuous is a perpetual mission that compacts its own context as it runs and self-heals through outages, so a long-running job doesn't fall over the first time a provider hiccups. For heavier setups, it can even fan work out across a fleet using git worktrees, which is a nod toward running several agent instances in parallel on the same project.

This is a tool for developers who are comfortable in a terminal and want an agent they actually control. If you've found hosted coding assistants too locked down, or you want to bring your own model and keep the whole thing on your machine, Wizard leans hard in that direction. The bring-any-model stance and local inference support make it appealing to people who care about cost, privacy, or simply not being tied to a single API. It rewards tinkering, since the extension model invites you to shape the agent around your own workflow rather than the other way around.

The project is honest about its limits, which is a good sign. Small local models are plainly worse than frontier models, so the bring-your-own-model freedom comes with the trade-off that a tiny local model won't match a hosted flagship. There's no execution sandbox, meaning tools run with your own privileges, so you're trusting the agent with real access to your machine. Windows users need WSL2 rather than a native build. It's also a young project with a small community so far, which shows in the star count, though it's under active development with frequent releases.

On access, Wizard is open source and free to install and run under its MIT license, and there's no paid tier for the software itself. Whatever you spend goes to the model providers you choose, or nothing at all if you run everything locally. For someone who wants a hackable, provider-agnostic agent that lives in the terminal and can grow new abilities on demand, it's an unusually ambitious take on what a personal coding agent can be, and the one-line install lowers the bar to trying it.