Built with Strands Agents

Disk pressure,
handled quietly.

QuietGuard detects runaway diagnostic logs, makes a bounded cleanup plan, and acts only inside explicit safety boundaries. Uncertain evidence reaches a human with context.

Watch the 2:04 demo Explore the code

Open source · MIT licensed · deterministic offline demo · no cloud account required

Cycle complete
QG-20260905-01
3.0 MB safely reclaimed in the contained incident
2allowlisted logs removed
2sensitive files preserved
1decision escalated
✓ Audit chain verified · exact root boundary held
Read-only by defaultActions require an explicit flag
Marker allowlistOnly approved subtrees qualify
Budget boundedEvery cycle has a byte limit
Evidence preservedSHA-256 hash-chained audit
Working proof

See the whole agent cycle.

The demo creates a contained log storm, runs the real Strands tool loop, applies guarded actions, and publishes the evidence. Captions explain every decision.

Agent loop

Evidence before action.

A real strands.Agent coordinates four narrow tools. Each step leaves a durable artifact that can be reviewed or replayed.

Scan

Walk the exact configured root, refuse links and reparse points, and compare files with the previous baseline.

Plan

Classify evidence by marker, extension, age, and risk. Fit safe candidates inside the cycle budget.

Act

Revalidate every boundary immediately before deletion. A changed or uncertain file is skipped.

Publish

Write the dashboard, machine-readable incident report, and tamper-evident audit chain.

QuietGuard dashboard showing 3.0 MB reclaimed, protected files, and one review escalation
Safety contract

Autonomy with a hard edge.

QuietGuard removes known regenerable noise and refuses to guess about user data.

✓
Explicit operator boundaryA .quietguard-safe marker is required below the exact configured root.
✓
Narrow file policyOnly old .log, .tmp, .trace, and .dmp files qualify.
✓
Unknown means protectedDatabases, documents, media, archives, unmarked logs, links, and reparse points remain untouched.
✓
Reviewable decisionsEach decision records its evidence in a SHA-256 linked audit stream.
Architecture

One agent. Four focused tools.

The included deterministic Strands model makes the full demo private and free to run. A production deployment can swap in any Strands-supported model without changing the filesystem guardrails.

QuietGuard architecture: disk signals flow into the Strands agent, scan, plan, guarded action, dashboard, and human review
Try it locally

Run the contained incident.

Python 3.10+ is enough. The demo creates its own synthetic workspace and never touches your files.

git clone https://github.com/PionneerZ/quietguard-agent.git
cd quietguard-agent
python -m pip install -e .
quietguard demo