Phanto ShieldComing soon
Phanto Shield
Security for the AI agents you run on your own computer, and for the memory they keep.
What Phanto Shield is building
See what your agents can reach
A plain list of the files, tools and keys each local agent has access to.
Protect agent memory
Watch the notes and memory your agents keep for tampering and for secrets that should not be there.
Keep secrets out of logs
Spot keys and passwords that end up in logs, prompts or saved memory.
Review what happened
A readable activity trail, so you can check what an agent did and when.
The blueprint: eight protections
A free, vendor-neutral blueprint for a security layer that sits around an AI agent you run yourself. It describes the ideas, not any one product's code.
Untrusted-input screening
Every inbound message, web page or document is treated as data, never as instructions. Suspicious phrasing is flagged before the agent sees it.
Tool-call approval gates
Risky actions such as sending, deleting, paying or publishing pause for a yes from you. Safe, read-only actions can run on their own.
Secret redaction and scanning
Keys, tokens and passwords are masked in prompts, logs and saved memory, and scanned for in files the agent can reach.
Kill switch
One control that stops the agent and cuts its connectors straight away, so you can act fast when something looks wrong.
Audit log
A readable, append-only record of what the agent was asked, what it tried and what you approved.
Memory write review
New long-term memories are checked, and anything odd is held for your review so one bad message cannot rewrite what the agent believes.
Least-privilege connectors
Each tool or account gets only the access it needs, with read-only as the default and short-lived permissions where possible.
Backup and restore drills
Regular backups of agent settings and memory, plus a practised restore, so a bad day is an inconvenience and not a loss.
How it will connect to your agents
Phanto Shield is planned as a local layer. The idea is that it runs on your own machine and your data stays with you.
Step 1
Install locally
A small local app or proxy that runs next to your agent. No account is needed for the free blueprint.
Step 2
Point your agent at it
Your agent talks to its tools and model through the layer, so inbound content and outbound actions pass a checkpoint.
Step 3
Choose your rules
Start from a safe preset. Decide which actions need approval, which connectors are read-only and what counts as a secret.
Step 4
Review and rehearse
Read the audit log, approve held items and run a restore drill on a schedule you set.
Planned integrations
Designed to sit in front of the kinds of tools people run at home or in a small business. Support will be listed here as each one is ready.
MCP clients
Agents and apps that use Model Context Protocol tool servers.
Local model runners
Desktop and self-hosted tools that run language models on your own computer.
Memory and vector stores
The notes, files and databases an agent uses to remember things.
Inbound channels
Email, chat and web content that can reach your agent from outside.
Phanto Shield is a planned product. Free blueprint: available as a guide when it is published. Paid: coming soon. Nothing is installable today, and no security tool can promise that an agent is risk-free.
Plans
Start free. Paid plans are not open yet and no prices are set.
Free
Free
- Check one local agent setup
- Basic secret and permission checks
- Account opens at launch
- Limits will be published before launch
Paid
Coming soon
- Cover more agents and machines
- Ongoing monitoring and alerts
- Longer activity history
- Pricing is not set yet
Get notified
Run AI agents on your own machine and want them locked down? Email or call us and we will add you. The sign-up form is coming soon.