by AceTeam — safety infrastructure for AI agents

SafeClaw.sh

Your AI agent just spent $135K. Deleted an inbox. Leaked a credit card number.
Turn safety on.

SafeClaw = OpenClaw (AI agent) + Agent Safety Net (local proxy). The agent runs in a container. The safety net blocks dangerous actions, tracks cost, and signs every decision.

Star on GitHub Fork
QR code to safeclaw.sh

Scan to get started

At a workshop or lecture? Point your phone at the QR code to open this page. Star the repo and connect on LinkedIn to keep this project moving and show some love!

Star on GitHub · Connect on LinkedIn

New to AI agents? This site works for everyone — students, first-time builders, and seasoned engineers. Follow along in order, or jump to the part you need.

Available now

Hosted

Your own SafeClaw instance on AceTeam. No monthly fees — just pay for token usage at fair market rate.

Launch on AceTeam
Free to start. You only pay for the tokens you use.
Available now

Self-Host

Your data never leaves your machine. $0 forever.

Setup Guide
Your keys, your hardware. Apache 2.0.
Without SafeClaw
  • No spending visibility
  • Agent can delete files, scan ports, run exploits
  • Personal data can leak to any API
  • No record of what happened
With SafeClaw
  • Cost per call in real time
  • Dangerous actions blocked before they happen
  • SSNs, credit cards, and PII caught and stopped
  • Every decision signed and auditable
Claude Code OpenClaw CrewAI LangChain NemoClaw NanoClaw Any LLM API

How It Works

SafeClaw sits between your agent and the LLM. Every call is checked before it reaches the model. Dangerous requests are blocked. Everything is logged.

Works with Claude Code

Once SafeClaw is running (self-hosted or via the installer above), add this to your Claude Code MCP config to give it 4 safety tools — cost lookup, verdict history, safety status, and threat scan:

{"mcpServers":{"aceteam":{"type":"streamable-http","url":"http://localhost:8899/mcp/"}}}

Paste into ~/.claude.json (or your project's .mcp.json). Or run aceteam-aep setup to auto-detect and configure.

Self-Host Setup

What gets installed? The installer gives you two things: OpenClaw (the AI agent platform) and the Agent Safety Net (blocks threats, tracks cost, signs verdicts). Together, that's SafeClaw. Your data stays on your machine — nothing touches our servers. You use your own API keys, so it costs $0 beyond what you already pay your LLM provider.

Have OpenClaw or another AI agent? Tell it to install SafeClaw for you:

I'm a Human
I'm an Agent
1
Send this to your agent (paste into OpenClaw, Claude Code, or any MCP-capable agent):
/skill https://safeclaw.sh/SKILL.md
Or: openclaw skills install safeclaw-proxy
2
Pick a provider — open the dashboard and click Use AceTeam. Sign in at aceteam.ai once, get bounced back with a key already wired up — $5 free credit included. Or bring your own OpenAI / Anthropic key / TokenRouter.
3
Share it — Tweet about SafeClaw to help others discover it.

You are an AI agent. Run these steps to install SafeClaw:

1
Fetch and execute the skill
curl -fsSL https://safeclaw.sh/SKILL.md
Follow the instructions in the SKILL.md to install and verify the proxy.
2
Point your human at the dashboard — tell them to open localhost:8899/dashboard/ and click Use AceTeam. One sign-in at aceteam.ai, bounced back with a key and $5 free credit — no key handling on your side.
3
Start learning — explore the dashboard and run the demo scenarios below.
Recommended — One-Line Installer

Run the installer to automatically detect your environment, pull images, and set up your workspace.

Linux / macOS
curl -fsSL https://safeclaw.sh/install.sh | bash
Windows (PowerShell)
iwr -useb https://safeclaw.sh/install.ps1 | iex
Manual setup — run the container yourself
Safety proxy only
docker run -p 8899:8899 ghcr.io/aceteam-ai/aep-proxy

This runs just the safety net. For the full agent, use docker compose (see below).

Full SafeClaw (OpenClaw + Safety)
docker compose -f docker-compose.yml -f docker-compose.safe.yml up

Runs the complete agent + safety proxy stack. Using Podman? Replace docker with podman.

Next — Open the dashboard & pick a provider

Open localhost:8899/dashboard/. On first run, a welcome card asks how you want to reach an LLM. Three options — all store the key locally in the proxy, nothing touches our servers:

Use AceTeam — recommended, one click

Click Use AceTeam → sign in at aceteam.ai → you're bounced back with a key already provisioned. Comes with $5 in free credit; top up from your account when it runs out.

I have API keys — bring your own OpenAI / Anthropic / local

Paste an sk-… key. Optional base URL for local Ollama, LM Studio, vLLM, etc. $0 extra cost — you pay the provider directly.

Use TokenRouter — via palebluedot.ai

One key for 100+ models across providers. Base URL prefilled — paste your TokenRouter key.

Want more than $5? Here's how to top up.

After signing in with Use AceTeam:

  1. Connect with me on LinkedIn and say hi
  2. Star SafeClaw on GitHub

I'll add $20 in tokens to your AceTeam account on top of the $5 welcome credit.

3 — Point your agent
export OPENAI_BASE_URL=http://localhost:8899/v1

Every LLM call from any tool (Claude Code, OpenClaw, CrewAI, LangChain) now routes through SafeClaw before reaching the provider.

Already have OpenClaw? Wire it to the safety net manually

If OpenClaw is already installed, point it at a running Agent Safety Net via its config TUI:

uv sync --extra proxy
openclaw config

Then walk through the menu:

  1. Select options to configure → Model
  2. Model/auth provider → Custom Provider
  3. API Base URL → http://localhost:8899/v1
  4. How do you want to provide this API key? → Paste API key now → paste your key
  5. Endpoint compatibility → whatever you want
  6. Model ID → whatever model you want (e.g. gpt-4o)

OpenClaw will now route every call through the safety net. Start a new session for the change to take effect.

4 — Try a prompt

Ask your agent anything. Then check the dashboard — you'll see the cost, safety verdict, and a full audit trail of every call.

Alternative: developer mode (git clone + uv)
git clone https://github.com/aceteam-ai/aceteam-aep.git
cd aceteam-aep
curl -LsSf https://astral.sh/uv/install.sh | sh
uv sync --extra proxy
uv run aceteam-aep proxy
Clone the source, install uv, sync proxy extras, and run from the repo. Best for hacking on the proxy itself — changes to the source take effect immediately.
Self-Host (available now)
  • Data stays on your machine
  • Your own API keys — $0 extra cost
  • Full dashboard at localhost
  • Works offline (once container is pulled)
Hosted on AceTeam
  • No install — launch in one click
  • Full SafeClaw chat UI in your browser
  • Build apps, run workflows via MCP tools
  • Free — pay only for tokens at fair market rate
In a workshop? Click here for setup instructions

Self-hosted: Your instructor will give you a URL:

export OPENAI_BASE_URL=http://INSTRUCTOR_URL:8899/v1

Replace INSTRUCTOR_URL with the address your instructor provides.

Host a SafeClaw Workshop — for instructors and organizers

Running a workshop or classroom session? Here's how to set up for your attendees.

Self-Host for the Room

Run the proxy on your machine, share with attendees
  1. Start the proxy:
    aceteam-aep proxy --port 8899 --host 0.0.0.0
  2. Share your IP or create a public tunnel:
    ngrok http 8899
  3. Attendees set one env var with your URL

Materials: Bootcamp guide, demo scenarios, slides

Hosted

Each student gets their own SafeClaw instance
  1. Students go to aceteam.ai and launch an instance
  2. Free — no setup, just pay for tokens used
  3. Connect on LinkedIn & star the repo for bulk workshop coupons

Pricing

Free (Self-Host)
  • Full safety proxy + dashboard
  • Built-in threat detection
  • Complete audit trail
  • Open source (Apache 2.0)
AceTeam Hosted
  • Your own SafeClaw instance in one click
  • Chat UI, app builder, workflow tools
  • Safety dashboard + cost tracking built in
  • Free — pay only for tokens at fair market rate
Troubleshooting

"podman" or "docker" not found: Install Podman or Docker Desktop.

Can't open localhost:8899: Make sure the run command is still active.

Slow first download: ~1 GB one-time, then starts in seconds.

Stay in the Loop

Get SafeClaw release notes, workshop dates, and new threat detectors by email. No spam, unsubscribe anytime.

Want hosted SafeClaw? Launch your own instance at aceteam.ai — free, pay only for token usage.

Learn More

GitHub Safety Net (AEP) PyPI AceTeam Platform
Pale Blue Dot × AceTeam AceTeam

Partner offer: $200 in AceTeam credits — 50 vouchers available.

palebluedot.ai (makers of TokenRouter) is partnered with us. To claim: sign up at aceteam.ai, star the repo, then DM Jason on LinkedIn — mention TokenRouter.

SafeClaw by AceTeam · Open source (Apache 2.0) · Contact

The Agent Safety Net is powered by AEP (Agentic Execution Protocol) — an open standard for governing every LLM call an agent makes: cost tracking, PII detection, policy enforcement, and signed audit trails.