ZERO-TRUST SECURITY FOR AI AGENTS

The firewall for
MCP and A2A agents.

Protect Claude Code, Cursor, Codex, Gemini CLI, and every AI agent from prompt injection, tool poisoning, runaway loops, and unsafe tool calls.

Install in 30 seconds View live demo

MIT licensed · Local first · One binary · No account required

<5msp99 overhead
0bytes leave your box
5agents auto-detected
250+tests · 6 platforms
Live visibility

See what your agents are actually doing.

Real-time visibility into every MCP and A2A tool call with live blocking, cost tracking, and policy enforcement.

Like htop for AI agents.

agentguard tail
$ agentguard tail
▋
Capabilities

Everything between your agent and disaster.

Six layers of inspection on every tool call — so a poisoned response, a runaway loop, or a leaked secret never reaches your agent.

⛨

Prompt Injection Defense

Scans inbound and outbound tool traffic before the model ever sees it. Heuristic scorer flags injection signals in under 100µs.

↻

Loop + Cost Protection

Stops infinite loops and enforces spend caps per session with a sliding-window circuit breaker.

◎

Trust Scores

Know which MCP servers are abandoned, unsafe, or vulnerable — before you connect them.

⟲

Replay Sessions

Replay agent sessions locally and test new rule packs without paying for LLM tokens again.

⚙

Policy Engine

Write YAML rules once — allow, block, flag, redact — and enforce them everywhere.

✦

Zero Config

One command install. Auto-detects Claude Code, Cursor, Codex, Gemini CLI, and Windsurf — backs up every config byte-for-byte before touching it.

Architecture

Sits between your agent and the world.

Claude Code · Cursor · Codex your AI agent
↓
AgentGuard Sidecar 5-stage inspection · transparent proxy
↓
GitHub · Slack · Notion · Filesystem · APIs your tools & data

Transparent proxy. Sub-5ms overhead. Nothing leaves your machine.

Install

Install in one command.

Detected: detecting… 
$ curl -fsSL https://agentguard.space/install | sh

macOS 11+. Installs to ~/.agentguard/bin/, verifies SHA-256, adds to PATH.

$ curl -fsSL https://agentguard.space/install | sh

Linux amd64 + arm64. Single static binary, no glibc requirement.

PS> iwr -useb https://agentguard.space/install.ps1 | iex

PowerShell 5+. Installs to %USERPROFILE%\.agentguard\bin\, appends user PATH. Open a fresh terminal afterwards.

$ curl -fsSL https://agentguard.space/install | sh

Git Bash, Cygwin, MSYS2, or WSL — anything with POSIX sh. Same installer as macOS/Linux.

$ go install github.com/nikegunn/agentguard/cmd/agentguard@latest

Requires Go 1.23+. Lands in $(go env GOBIN) or $GOPATH/bin.

Cosign-signed releases ↗
The easy way

Let your AI agent set it up.

Setting up AgentGuard is exactly the kind of thing your agent is good at — run a command, check a config, start a dashboard. Pick your agent, copy the prompt, paste it into the chat. It does the rest.

Paste this into Claude Code 

      


      

        100% client-side — the prompt is assembled in your browser, nothing is sent
        anywhere. The agent backs up your config before changing it, and
        agentguard uninstall restores everything byte-for-byte.
      

    

  







  

    In the wild
    
What early users are saying.

  

  

    

      
“Every AI agent should ship with this.”— Staff Engineer, Series-B fintech

      
“This catches things MCP servers completely miss.”— Platform lead, enterprise SaaS

      
“The missing firewall layer for AI tooling.”— Solo founder, AI startup

      
“Caught a base64-encoded prompt injection on day one.”— Security engineer

      
“Byte-identical uninstall got us through security review.”— Platform lead

    

  







  

    
    

      Your AI agent is already connected to the internet.

      Now give it a security layer.
    

    

      Install AgentGuard
      GitHub ↗
    

    
Built in Nepal by Nikhil Bhagat.