OPEN SOURCE

SafeClaw

AI agents that ask before they act.

Intercepts every action your AI agent tries to take and checks it against a safety policy before execution. Local gating with browser dashboard, approval workflows, and audit ledger.

$npx @authensor/safeclawCopy
WHAT IT DOES

Safety for every tool call

Every tool call goes through a policy check before it runs.

Risky actions require your explicit approval. Nothing runs without your say-so.

Works with Claude and OpenAI out of the box.

100% open-source. MIT licensed. Free forever.

HOW IT WORKS

Five steps, one line of code

1
Agent decides to act
2
SafeClaw intercepts the call
3
Checks your policy
4
Allows, denies, or asks you
5
Action only runs after clearance
What leaves your machine

Action metadata only, e.g. "write /tmp/output.txt"

What stays local

Your API keys, your files, your data. Always.

CAPABILITIES

Everything you need to ship safely

Browser Dashboard

Setup wizard, task runner, and approval center in one UI.

Policy Engine

Deny-by-default rules, visual editor, and dry-run simulation.

Audit Ledger

Append-only log with SHA-256 hash chain for every decision.

Risk Signals

Badges on suspicious commands: obfuscation, credential access, exfiltration, destructive ops.

Budget Controls

Spending limits, usage analytics, CSV/JSON export.

Mobile PWA

Swipe-to-approve from your phone. SMS and webhook notifications.

Container Mode

Sandboxed execution for untrusted workloads.

Offline Cache

Fail-closed safety. Denies by default when disconnected.

Integrations

Slack, Discord webhooks. Claude and OpenAI SDKs supported.

TRUST

Built to be audited

446
tests across 24 files
0
npm dependencies beyond the SDK
MIT
open-source license
Deny-by-default securityCSRF protectionReDoS protectionSecrets redactionRate limitingRead every line

One command to get started

Your browser opens. A wizard walks you through everything.

$npx @authensor/safeclawCopy