PUBLIC ALPHA DOCS

Use the supported path. Keep the gaps visible.

These short guides cover the public Windows build and the invite-only staging dashboard. They do not turn the alpha into a complete security product.

Install and update

Download install.ps1 from this website, review it, and run it. The script downloads the pinned Windows amd64 executable, verifies its SHA-256 checksum, installs a stable agentguard.exe, and adds its directory to your user PATH.

powershell -NoProfile -ExecutionPolicy Bypass -File .\install.ps1
cd C:\path\to\a\disposable\project
agentguard init
agentguard doctor --dir .

Updates remain human-controlled: run agentguard version --check, review the new installer, and replace the binary only after checksum verification. AgentGuard does not silently download or install updates.

Open the full install, checksum, update, and rollback guide →

Publish a redacted summary

Publishing is optional and explicit. It sends the current session's task summary, deterministic classification, plain explanation, and content-free evidence references to the staging service. It does not publish raw commands, source or file contents, diffs, absolute paths, credentials, transcripts, or arbitrary tool payloads.

After an operator has invited you, sign in to the staging dashboard, create a device token, and use the command shown there. The supported publish command targets the staging service explicitly:

agentguard session publish --url https://staging.jonnary.com

The CLI reports success or failure. A hosted outage does not change local policy decisions, local evidence, or local review.

Trust boundaries

The local alpha observes only documented Claude Code hooks; unsupported shell, wrapper, MCP, network, and agent actions may be invisible.
The scope result is a deterministic path-and-order signal, not semantic task understanding, compliance evidence, or proof that a change is correct.
An allow result bypasses Claude Code's native prompt only for a strict routine rule. Invalid hook input fails open to the native permission flow.
Published summaries are self-reported and may be incomplete. Dashboard reviewers remain responsible for every conclusion and action.
The public website uses no JavaScript, analytics, cookies, signup form, account creation, or billing flow.

AgentGuard is a working title and an open alpha experiment. It provides no security guarantee.