Open source · v1.0.0-beta

An AI agent
you can trust

SafeClaw intercepts every action your AI agent takes before it executes.
Nothing runs without passing through your policy engine first.
Install in 3 steps. No experience needed.

Action-level gating
Every action intercepted and evaluated before execution. Not monitoring — prevention.
Configurable policy engine
Dynamic rules with conditions, not static permission lists. Simulate policies before deploying.
Any AI provider
Works with Claude and OpenAI out of the box. Not locked to any single agent framework.
1

Download Node.js

Free tool that runs SafeClaw. Download, open installer, click "Next" until done.

Download Node.js
2

Open Terminal

It's already on your computer:

Cmd + Space, type Terminal, hit Enter
Win + R, type cmd, hit Enter
Ctrl + Alt + T
3

Paste & run

Copy this, paste into Terminal, hit Enter:

$ npx @authensor/safeclaw
Your browser opens automatically with a setup wizard.
Scroll to learn more

Intercept first. Execute second.

Other tools monitor what already happened. SafeClaw gates every action before it runs — your agent can't do anything you haven't approved.

Deny by default

Every action is blocked until you create a rule that allows it. Your agent starts locked down, not wide open.

Human-in-the-loop approvals

When your agent wants to do something risky, SafeClaw pauses and asks you. Approve or deny from your browser or phone.

Tamper-proof audit trail

Every action is logged with a cryptographic hash chain. Nobody — not even SafeClaw — can alter the records.

Your keys stay local

API keys are saved on your machine with encrypted file permissions. They're never sent to our servers or anyone else's.

Configurable policy engine

Set rules like "allow file reads, block code execution, require approval for network requests." All from the dashboard — no code needed.

Works with Claude & OpenAI

Use Anthropic's Claude or OpenAI's GPT models. SafeClaw wraps around either and adds the safety layer on top.

How it works

SafeClaw sits between your AI agent and the outside world. Every action is intercepted, evaluated against your policy, and gated before execution.

1

You give your agent a task

Type something like "Organize my project files" or "Write a Python script to clean this data" in the dashboard.

2

SafeClaw checks every action

Before the agent writes a file, runs code, or makes a network request, SafeClaw evaluates it against your policy rules.

3

You stay in control

Safe actions proceed automatically. Risky ones pause for your approval. Dangerous ones are blocked entirely. You set the boundaries.

What you can automate

Give your agent real tasks and let SafeClaw handle the guardrails.

File management Organize, rename, and clean up project files across your machine.
Code generation Write scripts, fix bugs, refactor code — with every change reviewed.
Data processing Parse CSVs, transform datasets, generate reports automatically.
Scheduled tasks Run agents on a schedule — daily backups, weekly reports, continuous monitoring.
Research & analysis Let your agent read documentation, summarize findings, and draft responses.
System administration Manage configs, check logs, run diagnostics — all within your safety policy.

15 creative ideas

Things you didn't know an agent could do

15 useful ideas

Practical tasks that save real time every week

Why trust SafeClaw?

We built SafeClaw to be the most transparent, verifiable agent safety tool available.

100%

Open source. Read every line of code on GitHub.

418

Automated tests. Security-hardened and continuously verified.

0

Third-party dependencies (beyond the AI SDK). Minimal attack surface.

Ready to try it?

One command. Your browser opens. The wizard walks you through everything.

$ npx @authensor/safeclaw
Star on GitHub