A product under RAVIONIX Coming soon
RAVIONIX AI Guard See what the AI agents on your PC can do.
AI tools like GitHub Copilot, Cursor, Codex, Gemini and Claude can run commands and read files on your PC. RAVIONIX AI Guard shows what they’re allowed to do and explains the risks in plain English. It’s designed to do this without uploading anything.
Coming soon. AI Guard is in development and not released yet, so you can’t download it. There’s no release date or pricing yet.
What AI Guard does
It finds the AI agents on your PC, shows how much each one can do on its own, and helps you make them safer.
Your PC · offline
Your AI agents
- Gemini CLI
- Codex
- Claude
- GitHub Copilot
- Cursor
- Windsurf
- + more
Settings read only. Never run, chats never read.
RAVIONIX AI Guard
- 01FindAgents, exposed keys, risky add-ons, fake AI apps
- 02RateFrom “Asks first” to “Full control”
- 03FixOne click, with your OK
- 04WatchAlerts when an agent gets more freedom
You
See it in plain English
Every finding says where it is, why it matters and what to do.
- Preview
- Apply
- Undo
Nothing changes until you say yes.
What it checks in each agent
Pick an agent to see the settings AI Guard looks at.
What it checks in every agent
For every supported agent: whether it asks before acting, what it can run without asking, which tool servers and add-ons it loads, and any keys left where it can read them.
Approval settings · allowed commands · MCP servers · add-ons · exposed keys
What it checks in Gemini CLI
Whether its approve-everything mode is locked off, and which tool servers and tools it trusts without asking.
YOLO lock · trusted servers · allowed tools
What it checks in Codex
Whether it asks before acting, how much of your PC it can reach, and which projects it trusts.
Approval policy · sandbox mode · trusted projects · MCP servers
What it checks in Claude
In Claude Code and Claude Desktop: whether it asks before acting, what it may run without asking, and the hooks, tool servers, skills and plugins it loads.
Permission mode · allow rules · hooks · MCP servers · skills, plugins and extensions
What it checks in GitHub Copilot
In VS Code and the Copilot CLI: auto-approve settings that let it act without asking, and the tool servers it loads.
Auto-approve · terminal approvals · MCP servers
What it checks in Cursor
Whether it runs commands without asking, what it allows, and whether its sandbox is turned off.
Auto-run · command allowlists · sandbox · MCP servers
What it checks in More agents
Windsurf, Cline, Roo Code, Continue, OpenClaw, LM Studio and JetBrains AI: the tool servers they load, what runs without asking and, for OpenClaw, who can message it and whether it can run commands. Other AI tools show up as “found, not checked yet” instead of staying invisible.
MCP servers · auto-approve settings · messaging channels · start-up entries
Fix risky permissions
Only when you clickWhen Claude Code is allowed to do too much, a “Fix this” button lets you change that setting in one click, and “Harden settings” adds safer defaults in one go. For example, you can:
- Make Claude Code ask you before it acts
- Keep Claude Code's bypass-mode warning on
- Turn off a plugin that runs commands automatically
- Turn off a plugin or tool server a project switches on
- Block access to SSH keys, AWS credentials and .env files
Before
Claude Code is in bypass mode and acts without asking you.
After
Claude Code asks you before it acts.
Preview
See the exact file and setting that would change.
Apply
Only when you click. Only that one setting changes.
Undo
Every fix is listed, so you can undo it any time.
For everything else, including Copilot, Cursor and the other agents, each finding comes with numbered steps to fix it yourself.
It never touches settings your organisation manages, your chat history or saved sign-ins. If a change isn’t safe, it changes nothing.
More than a one-off check
Watch mode
AI Guard can stay in the notification area and tell you when an agent gets more freedom: its level goes up, a token is saved in a rule, a rule allows every command, bypass mode is turned on, or an agent starts with full control. It never changes anything by itself.
Keys an agent can read
Finds keys an AI agent could read outside its settings: shell history, the sign-in files of git, npm, AWS, Docker, GitHub CLI, kubectl, Wrangler, Vercel, Azure and gcloud, and SSH keys with no passphrase. It shows the kind and count, never the value.
Fake AI apps
Checks shortcuts, installed programs and Store apps that use an AI brand name, including lookalikes, against the real publisher’s signature, offline.
Your Windows account
Shows whether you’re an administrator, how strict UAC is, and any agent running as administrator, with step-by-step help to run agents in a separate standard account.
Protects itself
Anything that makes AI Guard less watchful, such as undoing a fix or turning Watch off, asks for Windows Hello first. Its own records are signed, so a change made outside the app is reported.
A report you can keep
Every scan ends with a plain-English report, including what the scan can’t see. You can export it as a self-contained HTML file.
How it’s designed
- Designed to work offline, with no uploads and no telemetry.
- Designed never to read your AI conversations or chat history.
- Designed never to show or save a full secret.
- Designed never to change a setting without your OK, with every fix previewed first and easy to undo.
- Designed never to run the tools it inspects.
AI Guard isn’t released yet. Once it ships, this page will describe what the released version does.
What a check can’t see
- Watch mode notices when an agent gets more freedom, but it can't see inside a tool you've already allowed, so a leak through that tool can look like normal work.
- It can tell you if it was stopped, but it can't stop a program that runs as you.
- It checks the AI tools it knows about. Others show up as “found, not checked yet”, and more are added over time.
AI Guard shows you what your agents can do. It can’t promise to stop every attack.