Scoot

English | 中文

Scoot is a small, safe AI agent that lives in your terminal.

You give it a goal. It thinks in steps, runs local tools to make progress, and writes down everything it did. No app to install, no cloud account, no state leaving your machine — just one binary and a model backend you choose.

scoot -e "find every TODO in this repo and summarize them by file"

What makes Scoot different

Most coding agents are large applications that trust the model and reach for the cloud. Scoot is the opposite by design.

• One tiny binary, no runtime. Written in pure Zig, it ships as a single self-contained executable. Copy it to a laptop, a NAS, an edge device, or a container and it just runs. (why Zig)
• Safe by default. Scoot never executes raw model output. Every step is validated, and every tool call passes through a policy gate that can block dangerous commands or refuse writes and network access entirely.
• Fully auditable. Every thought, tool call, observation, and decision is saved as plain JSONL. You can replay exactly what the agent did, after the fact.
• Local-first. Config, sessions, skills, and logs all live under ~/.scoot. Nothing syncs anywhere you didn't ask it to.
• Bring your own model. Any OpenAI-compatible backend works — local (Ollama, vLLM) or hosted (OpenAI). No provider lock-in.
• Extend without rebuilding. Drop a folder with a SKILL.md into your skills directory and the agent can discover and use it. (skills)

How it works

Scoot runs a ReACT loop. Each turn, the model returns one structured step, Scoot checks it, runs it, and feeds the result back:

flowchart TD
    A[Your goal] --> B[Model proposes one step]
    B --> C{Valid and allowed?}
    C -- no --> B
    C -- yes --> D[Run the tool]
    D --> E[Record the result to the audit log]
    E --> F{Done?}
    F -- no --> B
    F -- yes --> G[Final answer]

Loading

The model can only ask for a fixed set of built-in actions — read and edit files, search code, run bounded shell commands, make a single HTTP request, call a skill, and a few more. It can never invent a capability that bypasses the gate. See Built-in Tools for the full list.

Quick start

1. Install. One line installs the latest release for your platform:

curl -fsSL https://raw.githubusercontent.com/scootship/scoot/main/install.sh | sh

On macOS you can use Homebrew instead:

brew install scootship/tap/scoot

Prefer Docker, apt (scoot, scoot-wasm, and scoot-edge are all published there), or compiling a smaller build from source? See Installation.

2. Configure. The wizard creates ~/.scoot and writes your config:

scoot setup

It asks for your model backend and where to find the API token. Every config key is documented in Configuration.

3. Run a goal.

scoot -e "summarize this repository"   # one shot, prints the answer
scoot                                  # interactive REPL

Add --trace to watch the agent think and act in real time.

Staying safe

Scoot has three policy modes. Pick the one that matches how much you trust the task:

Mode	Use it for	What it does
guarded (default)	Everyday interactive work	Allows normal work, blocks catastrophic commands
readonly	Untrusted or unattended jobs	No writes, no shell, no network — reads only
unrestricted	Tasks you fully trust	No limits, still fully audited

guarded is a convenience tripwire, not a sandbox. For unattended jobs Scoot automatically drops to readonly — scheduled/daemon jobs coerce their mode, and a one-shot scoot --unattended -e "<goal>" clamps in-child to the local edge.max_job_policy ceiling (default readonly) so the command line can only lower policy, never raise it. Pair it with OS-level isolation when you need strong containment. Full threat model: Execution Policy & Security.

Run it unattended

Scoot can also run on a schedule as a foreground daemon — handy for periodic read-only checks, reports, or probes that a supervisor like systemd keeps alive:

scoot daemon run

Scheduling, triggers (every, at, cron), and daemon lifecycle are covered in Scheduling & Daemon.

Documentation

The full, bilingual user guide is the mdBook under book/:

• Installation — build, install, Docker, backends
• Design Philosophy — goals, non-goals, boundaries
• CLI Reference — every command and flag
• Built-in Tools — the agent's action set
• Execution Policy & Security — modes and threat model
• Skills — authoring and using skills
• Scheduling & Daemon — unattended jobs
• Sessions & Audit — local state formats
• Embedding API — the stable Zig package surface

Chinese chapters live under book/zh/src/. Project shape and intent are in the Roadmap (中文); contributor guidance is in AGENT.md (中文).

License

MIT. See LICENSE.