🏆 micro1 Agentic Workflows Hackathon 2026

The Intelligent Firewall for AI Agents

Autonomous agents can run shell commands, alter files, and leak proprietary secrets. AgentGate intercepts every tool call in real time — evaluating semantic threat intent with Tier 1 LLM classifiers, scrubbing secrets, and enforcing human-in-the-loop oversight.

Open Live Dashboard Quickstart Guide GitHub Repo ↗

One command launch: agentgate -config configs/agentgate.yaml -serve :8700

Tiered Architecture

Four Layers of Autonomous Defense

Every tool call clears multi-tiered verification before reaching your system.

🧠

Tier 1 LLM Risk Classifier

Powered by Claude Haiku with structured tool choice. Catches novel semantic threats like find / -delete and chmod -R 000 that bypass regex.

🛡️

Policy Engine (Tier 0)

Deny-by-default whitelisting, per-tool rules, and hard regex safety nets that block destructive commands in sub-millisecond execution time.

🕶️

Deep Secret Redaction

API keys, passwords, and tokens are scrubbed to *** in nested JSON arguments and tool replies before logs or agents see them.

👥

Human-in-the-Loop Checkpoint

Gray-zone calls in the [0.4, 0.8) score range prompt human reviewers for approval before any irreversible action executes.

Deployment

Start Securing Agents in Seconds

Wraps seamlessly around any Model Context Protocol (MCP) tool server.

Terminal Commands

# 1. Clone and build AgentGate
git clone https://github.com/gunjitdhakar15/AgentGate.git
cd AgentGate
go build -o agentgate ./cmd/agentgate

# 2. Run firewall with live web dashboard
./agentgate -config configs/agentgate.yaml -serve :8700

# 3. Run the 20-case adversarial evaluation
go run ./cmd/eval-harness -mock-judge

configs/agentgate.yaml

policy:
  tool_rules:
    - apply_to: "*"
      deny: true
    - apply_to: "read_file"
      allow: true
    - apply_to: "shell"
      allow: true
      arg_deny_pattern: "(rm -rf|del /s|shutdown)"
  redact:
    - keys: ["api_key", "password", "token"]
      replacement: "***"