Wizard

Wizard is an open-source, self-hosted terminal agent written in Rust that runs inside a TUI and executes tasks autonomously on your machine. The workflow is direct: launch the binary, connect it to your LLM provider of choice, and issue a goal. The agent breaks the goal into steps, runs terminal commands, observes the output, and continues — branching based on what each step returns rather than waiting for you to intervene at each command.

The differentiating claim is self-extension: the agent can modify or expand its own behavior mid-task, meaning it is not strictly bound to a fixed toolset defined at startup. This puts it in a category of agents that run tasks on their own and adapt as those tasks evolve, rather than executing a static instruction list. For developers experimenting with autonomous agent architectures on local hardware, this is the feature that separates Wizard from a scripted LLM wrapper.

Wizard fits best in developer and research environments where you control the machine, you trust the model, and you want to reduce the manual overhead of running multi-step terminal workflows. It breaks down in contexts that require detailed logs of every issued command, approval before destructive operations, or integration into a CI/CD pipeline with structured output. The documentation surfaced on the project page is sparse — which means the gap between ‘agent does something unexpected’ and ‘I understand why’ is filled by reading the source, not a troubleshooting guide.

Installation is via a provided script and the binary is self-contained, so there is no runtime dependency chain to manage. Multi-provider LLM integration is handled at configuration time, letting you switch between a local model and a cloud endpoint without restructuring your workflow. No API of its own is exposed, so Wizard is a standalone tool — not a component you embed into a larger service.