VibeArmory
The VibeArmory guard tower: a stone watchtower with lit green windows and an open green-lit door.

See everything your AI agents do on your computer.

Monitor every agent, every file, every network flow. Block what you don't want.

Get VibeArmory v0.2.0 (Mac) See the features
  • All data stays on your computer
  • No account, no cloud
  • Secrets masked by default
  • Native app and CLI

Why

Agents are running freely on your computer. That's a big security and privacy risk.

Everything it sees is rated for risk: each connection, each agent session, each system item and each sensitive file.

01 Network Flows

Which app talks to which host, and what it sends.

02 Agents

Every command, file and connection, per session.

03 System

Launch items, extensions, ports and sensitive files, rated.

04 Files

Every file open in any app that uses the network, and every file an agent reads or writes.

Features

One app for network, file access, and agent flows.

Every connection, live, per app

Each flow and the program behind it, rated for risk as it happens.

  • Block, mark known or skip decryption from the row.
  • The same view in a terminal: watchdog live.

Block by domain, IP or URL

For every app or just one. Domain and IP rules hold on every port.

  • Review new destinations, with their risk reasons and recent requests.

Blocking is per request on inspected traffic and in agent sessions, and per connection on traffic you don't decrypt.

Inspect HTTPS traffic, secrets masked

Opt in, trust the local certificate authority, and read each request's headers and body.

  • Tokens, keys and cookies masked; unmasking takes Touch ID.
  • Export as HAR, JSON or cURL.
  • Skip any domain or app; apps that pin certificates are bypassed automatically.

Every agent session, step by step

Each Claude Code and Codex session, with every file it touched and every connection it made, rated for risk.

  • Tree, Map, Folders and Network views.
  • Block a destination straight from a session.

A file fence for AI agents

Keep each session inside its project and the folders you allow.

  • Alert records, Ask asks you, Enforce refuses.
  • Answer from a banner, a notification or watchdog fence-asks.

Needs the Developer ID build, pending Apple's Endpoint Security entitlement. Coming

Risk ratings and Insights

Every row is rated High, Medium or Low with reasons. Patterns become findings with a recommendation.

  • Large uploads, credentials over plain HTTP, unsigned programs, exposed services.
  • Computed on your computer by fixed rules. No reputation feed.

All rules in one place, broadest first

Network and file rules together, so the widest ones are easy to question.

  • Move them between Macs with watchdog rules.

Visibility

What's installed, what's listening, all data flows. Full visibility.

Background Apps

Launch items

Every launchd job, rated. A changed item comes back for review.

Extensions

System extensions

System and kernel extensions, VPNs and content filters.

Ports

Listening ports

Every listener, and whether your network can reach it.

Files

Sensitive files

Credentials and wallets held open by networked programs.

Processes

Programs on the network

Each one's signer, and everything it did.

Data Transfer

Bytes in and out

Per program and destination, over any window.

CLI

Everything, from the terminal

watchdog talks to the same agent as the app, with --json for scripts.

  • watchdog live, agents, insights, ports
  • watchdog run --ask -- claude: one fenced command.
$ watchdog domains --risk high
DESTINATION                 PROCESS   RISK   OUT
upload.example-paste.net    curl      high   2.1 MB
tunnel.example.run          python3   high   14 KB

$ watchdog block upload.example-paste.net
Blocked upload.example-paste.net for every program.

$ watchdog run --ask -- claude
Fenced in ask mode: ~/code/acme-web
ask  read ~/code/shared-config (outside the session's folders)

$ watchdog fence-asks allow 1 --project
Allowed ~/code/shared-config for project acme-web.
The CLI, illustrated with sample output.

How it works

Four native parts, one record on your computer.

Network firewall

Sees each flow and its program, and enforces domain and IP rules.

Network Extension framework

File firewall

Checks a Claude Code or Codex session's file access against the fence.

Endpoint Security · Developer ID build

Background agent

Stores, decrypts, rates and analyses.

Runs as you, in the background

App and CLI

Show the record and carry your decisions back. Changes from inside an agent session are refused.

SwiftUI app · watchdog

Privacy

Your traffic stays on your computer.

No cloud service, no account.

Local only

Everything is kept on your computer. Its only lookups are reverse DNS for addresses it asks you to review.

Encrypted at rest

Request contents and full URLs are encrypted with a key in your Keychain.

Masked by default

Secrets are hidden everywhere. Revealing them takes Touch ID.

Kept a while, not forever

Contents for 3 days, connections for 30, by default.

FAQ

Questions.

Which agents does it understand?

Claude Code and Codex. The file fence covers both.

Do I have to decrypt HTTPS?

No. You still get per-app connections, domain and IP rules, and agent tracking. Inspection adds request contents and URL rules.

Can an agent turn VibeArmory off?

Changes to rules, settings and the fence are refused when they come from inside an agent session. In Ask and Enforce, the fence also blocks a session's signals to VibeArmory.

Does anything leave my Mac?

No recorded data. No cloud, no account, no analytics.

How much does it cost?

It's in early access. Get VibeArmory.

Run your agents with the lights on.

Early access for developers on macOS.

Get VibeArmory v0.2.0 (Mac)