In development · controlled previewThe platform guard is in development. Public installation and production support are not yet available.
Policies, machine activity, and approval workflows in Shielda's controlled preview.
A separate TypeScript library you can install today. Try the published SDK
01 / Protect your team's agents
Give your assistants
some ground rules.
Platform Agent Guard is being built to put rules around coding assistants and the tools connected to them. Your team can see sensitive actions and decide which ones to allow.
This platform guard is developed separately from the public Agent Guard SDK. The open-source SDK is available today with its own scope and release status.
Know what your assistants can reach
See which assistants and connected tools are configured. Notice when a tool’s description or inputs change, so a familiar name doesn’t hide a different action.
Set the rules for risky actions
Set rules for commands, sensitive files, and where an assistant can send information. Watch activity or enforce the supported rules. Bring actions that need a person to their attention.
Catch secrets before they leave
Check supported outgoing tool calls and prompts for credentials and other configured secret values. Record which rule matched without copying the secret into the activity log.
Keep control across the team
Bring machine activity, approvals, and policy into the dashboard. Signed rules, a stop switch, and background checks help keep the guard configured as intended.
Claude Code · Cursor · Codex · Copilot · Windsurf · Gemini CLI · Cline
Local MCP relay and Claude Desktop extension wrapping
Coverage depends on the host and installation. Some hosts allow actions when a hook fails; Claude Desktop coverage is limited to the MCP paths wrapped by the guard. Managed settings and sandboxing are needed for stronger enforcement. We verify the exact setup during a preview.
Coming next / Access you can understand
See who gave the agent access.
And what it can pass on.
We’re building a map of who gave an agent access and what it can pass to another agent. Some inventory and export work exists in development; the interactive view and cloud connections are still planned. The map should keep conditions, gaps, and old evidence visible.
The investigator has read access to this repository for this task. The request fits both the parent grant and the narrower delegation.
This concept shows the intended access model. Delegation can narrow a grant; it does not automatically combine permissions. The map displays evidence and unknowns; execution still requires its own authorization.
02 / Safeguards inside Shielda
Our agents have
to ask, too.
Shielda’s own agents help investigate problems, review designs, and prepare fixes. They need permission for a particular job, review for proposed changes, and limits on what they can run.
Authority for a specific job
Purpose-bound AI requests, capability-scoped credentials, approval records, and budgets connect access to the task being performed.
Controlled execution
Scanner plans are signed and checked locally. Tool execution is separated from the control agent, with restricted containers, bounded resources, and limited network access.
Evidence you can inspect
Untrusted context is sanitized, sensitive output is redacted, and source references are rechecked. Proposed actions and observed results keep their own records.
These controls are implemented in the development codebase. Scanner execution and direct provider access remain restricted while release work continues. The controls described here are not a claim of independently verified production protection.
The next steps
From a developer's machine
to a company's policy.
Current development focuses on managed enrollment, device health, broader host verification, signed releases, and approval workflows. The permission map will make access and delegation easier to inspect as the product grows.
Our open-source projects are a separate way to explore Shielda's work today. Their repositories describe what is available and how to use it.
Start with your workflow
Which agents are
working in your company?
Tell us what they can access and where you want a person in the loop.