Local egress firewall for AI agents
Your agents can delete repos.
AgentG stops them.
AgentG watches every outbound call from your AI agents, developer CLIs, and MCP tools, holds the risky ones for your approval, and records a tamper-evident audit trail. All on your machine, no hosted control plane.
curl -fsSL https://app.agentg.dev/install.sh | shFree 7-day trial. No card required. macOS-first, Linux core.
Install
One static Go binary. Homebrew, install script, or a signed release tarball.
Activate
Sign up, copy your license token, and run agentg license activate <token>. Verified offline.
Protect
agentg setup installs the runtime, proves protection on your machine, and opens the dashboard.
Govern what leaves your machine
Exact per-agent identity
Signed launch context ties every request to the exact OpenClaw agent or Hermes profile that made it, not just the process name.
Real request decoding
REST, GraphQL, and MCP tool calls are decoded into concrete actions, so a repo read passes while a repo delete is held.
Human approvals
Risky calls pause for a decision in the dashboard, Slack, or a signed approval link.
Tamper-evident audit
A hash-chained SQLite audit log you can verify with agentg verify-audit.
Offline license verification
Your license is an Ed25519-signed token verified locally. No network needed, no phone-home to keep working.
Fail-safe by design
Recovery paths stay unlicensed and a lapsed license never breaks your network: protection degrades to observe-only.