Skillget

wizard

mcp
Security Audit
Fail
Health Warn
  • License — License: MIT
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Low visibility — Only 5 GitHub stars
Code Fail
  • rm -rf — Recursive force deletion command in install.sh
Permissions Pass
  • Permissions — No dangerous permissions requested

No AI report is available for this listing yet.

SUMMARY

Self-extending autonomous agent in one Rust binary. One-line install, any provider (OpenAI-compatible, Anthropic, xAI) or fully local via llama.cpp, live /evolve self-modification, MCP, messaging gateway, built-in bench

README.md

Wizard

CI

One line. Your sovereign agent. Self-extending. Bring any model.

Wizard fixing a bug: provider list, prompt, live edit, diff

curl -fsSL https://raw.githubusercontent.com/teddytennant/wizard/main/install.sh | bash

One command installs the wizard binary and a small default loadout (a Playwright browser over MCP and four subagents). The first run asks which provider you want and sets up the rest. Pick Local and Wizard sizes a Qwen 3 GGUF to your hardware and runs llama.cpp's llama-server itself, no API key needed. Or bring a key for OpenAI, Anthropic, xAI, OpenRouter, Cloudflare Workers AI, or any OpenAI-compatible endpoint, and switch live with /provider. It's one fast Rust binary on Linux and macOS; everything it learns is plain TOML under ~/.wizard/ that you can edit or delete.

Other ways to install: local-stack preinstall, minimal, bring-your-own-model, Nix, macOS, plus a first-run walkthrough, all in Getting started.

What it does

  • Any model, switchable live. Speaks the OpenAI-compatible chat API (streaming + native tool calls, with a prompt-based JSON fallback), so OpenAI, Groq, vLLM, LM Studio, OpenRouter, Cloudflare Workers AI, Anthropic, xAI, and Ollama all work. /provider switches the live agent between them; keys live in env vars or ~/.wizard/credentials.toml (mode 0600), never in plaintext config. → Providers
  • Runs models locally, fully managed. Pick Local and Wizard downloads a GGUF sized to your VRAM, then starts, supervises, and reuses llama-server for you, including a Metal build on Apple Silicon. → Model tiers · Bring your own model
  • Model fusion (/fusion). Run a panel of your providers as a debate and synthesize one tool-capable answer that tends to beat the best single model in the panel. → Fusion
  • Self-extension (/evolve). Add skills, MCP servers, scripted tools, and subagents as plain files that go live on /reload. Gated by a clean cargo build and a smoke test, it can also rebuild its own binary. Every change is logged, and the prior binary is kept one mv from rollback. → Self-extension
  • Runtime MCP. stdio and HTTP MCP servers merge into the tool registry without a rebuild: the path for computer use, browser control, and databases. → Self-extension
  • Genie / Sovereign modes, plus --continuous. An interactive direct-action TUI, a headless self-directing mode, or a perpetual mission that compacts its own context and self-heals through outages. → Modes
  • wizard bench. Records your real tasks as trajectories and replays them in isolated git worktrees to score builds and models against each other. "The new model is better" becomes a number. → Bench
  • Messaging gateway. Run headless as a bot you talk to from your phone (Telegram), each inbound message a sovereign agent turn in your project. → Gateway
  • Make it your own. After a deep evolve modifies its source, /publish forks upstream to your GitHub and hands out a one-line installer for your variant. → Fork and distribute

Smaller attack surface by construction. A single memory-safe Rust binary: no interpreter to inject into, no garbage-collected runtime. Every install also converges on a different /evolve loadout, so there's no uniform tool surface to target. Read SECURITY.md before autonomous runs; tools execute with your privileges.

Limitations

  • Platforms. Linux (x86_64, aarch64) and macOS (Apple Silicon and Intel). Windows isn't supported; run it under WSL2. The installer downloads a prebuilt binary for your platform and builds from source when one isn't published yet.
  • Small local models are worse than frontier models. A 9B–36B quantized Qwen will misformat tool calls, miss context, and need more steering than Claude- or GPT-class models. Wizard mitigates with native tool-call probing, a JSON fallback, and retry prompts; the 27B+ tiers make much better agents than the 9B tier.
  • No sandbox. Tools run with your privileges, with no per-action approval gate in either mode. Read SECURITY.md before running on anything you don't trust, and prefer a container/VM for autonomous or continuous work.
  • Context windows are finite. Wizard searches and reads selectively rather than ingesting the whole repo, but long sessions eventually push out early context.

Docs

Development

Rust 2024, Ratatui, Tokio. Single binary, < 60 MB stripped.

git clone https://github.com/teddytennant/wizard
cd wizard
cargo build --release
./target/release/wizard

There's also a Nix flake: nix run github:teddytennant/wizard, or nix develop for a shell with the Rust toolchain and llama-cpp.

Acknowledgements

Local inference is powered by llama.cpp (ggml-org): Wizard installs its llama-server when you pick the local option. Ollama is a first-class supported provider.

License

MIT; see LICENSE.

Author

Teddy Tennant (github.com/teddytennant)

Reviews (0)

No results found