AgentGate — Firewall for AI Coding Agents

AgentGate sits between AI coding agents (Claude Code, Cursor, Continue.dev, Aider, Gemini CLI, OpenAI Codex) and your machine. It intercepts every Bash / Read / Edit / Write call, applies a policy, scans outgoing HTTP for data-exfiltration patterns, and asks a human via Telegram / Discord / Slack when the call is risky.

Why

Install

pip install agentgate-firewall
agentgate doctor        # check pre-requisites
agentgate init          # write a starter policy.yaml
agentgate start         # launch the proxy on :18790

Then install the Claude Code / Cursor / Aider hook:

agentgate install-hook           # Claude Code
agentgate install-cursor-hook    # Cursor
agentgate install-continue-hook   # Continue.dev
agentgate install-aider-hook      # Aider
agentgate install-gemini-hook     # Gemini CLI
agentgate install-codex-hook      # OpenAI Codex CLI

Policy example

version: 1
default: deny

rules:
  # Read access to dotfiles is OK
  - id: allow-read-configs
    match: {tool: Read, file_glob: "~/.config/*"}
    action: allow

  # Block destructive Bash commands
  - id: deny-rm-rf
    match: {tool: Bash, command_regex: "rm\\s+-[rf]+.*"}
    action: deny
    reason: "Recursive delete blocked"

  # Allow `kubectl apply` but require human approval
  - id: ask-kubectl-apply
    match: {tool: Bash, command_regex: "kubectl apply.*"}
    action: ask

  # CEL-lite `when` conditions
  - id: deny-rm-elsewhere
    match: {tool: Bash, command_regex: "rm -rf.*"}
    action: deny
    when: 'event.cwd != "/srv"'

  # Rate-limit noisy rules
  - id: ask-deploy
    match: {tool: Bash, command_regex: "kubectl apply.*"}
    action: ask
    rate_limit: {capacity: 5, refill_per_sec: 0.1}

network:
  allowed_domains:
    - "*.github.com"
    - "*.anthropic.com"
    - "registry.npmjs.org"
  require_https: true

Subcommands

Security features

• Bash/Read/Edit/Write intercept — every tool call passes through policy.evaluate().
• CEL-lite when — gate matches on event fields (event.cwd != "/srv").
• Token-bucket rate limiting — per-rule throttling without bricking the agent.
• DLP egress scan — 50+ API-key patterns, JWTs, private keys, DB connection strings.
• Prompt-injection scanner — 28+ markers (ignore previous instructions, role overrides, etc.).
• Network allow/block lists — domain + HTTPS enforcement + per-rule exemptions.
• Tamper-evident audit chain — SHA-256 chained rows; integrity verifiable.
• Ed25519-signed receipts — optional cryptographic proof of audit entry provenance.
• HMAC-signed webhooks — verify callbacks actually came from AgentGate.
• Multi-channel approval — Telegram / Discord / Slack / approval server / file.
• Multi-environment manager — dev / staging / prod switching.

Supported agents

Contributing

See CONTRIBUTING.md. Run agentgate doctor first to verify your environment.

License

Apache 2.0 — see LICENSE.