Agent Safety Kit

English · 한국어 · 中文

Safety rails for people running local AI agents, coding agents, and autonomous workflows.

Agent Safety Kit helps catch the boring failures that break real agent setups: unsafe file mutations, unreviewed community skills, missing handoff context, and secret-like values in operational reports.

It is designed for operators who want local-first guardrails before they give an agent more tools, more autonomy, or more access.

The public core is intentionally local-first:

• no external sends by default
• no secret values in reports
• no runtime mutation unless a command explicitly performs a local file action
• JSON reports suitable for CI, cron, or human review

Why This Exists

Local agents are getting useful enough that people now let them edit files, run scripts, inspect repos, and coordinate longer tasks. The risk is not only model quality. It is operational drift:

• a tool deletes or overwrites files without a rollback path
• a community skill asks for more access than expected
• a long task gets interrupted and the next session cannot reconstruct state
• logs accidentally include secrets or private data

Agent Safety Kit makes those failure modes visible before they become expensive.

Free Core CLI

File Guardian Lite

Wrap risky file operations with an action log, snapshots, managed trash, and rollback metadata.

askit file write ./scratch/example.txt "hello" --reason "demo"
askit file delete ./scratch/example.txt --reason "demo delete"
askit file rollback ACTION_ID --reason "undo demo"

Skill Intake Gate

Classify community agent skills before installing them or granting access.

askit skill-intake examples/skills.json --out reports/skill-intake.json

Decisions:

• safe-to-review
• needs-approval
• audit-before-use
• blocked

Session Watchdog

Create a compact checkpoint before long-running work, model context pressure, or handoffs.

askit checkpoint --config examples/session-watchdog.json --out reports/checkpoint.json

Sponsor Pro

Sponsor Pro is a private repository for teams and power users who want deeper agent safety checks without exposing their private agent system.

• Context Pulse Lite
• TaskFlow Audit
• Memory Hygiene Probe
• dashboard templates
• CI examples
• adapter templates

See docs/pricing.md, docs/sponsor-tiers.md, and pro/README.md.

Agent Safety Audit

Need a one-time review of an agent repo, workflow, or tool setup? The launch audit checks for unsafe skill metadata, secret-like report content, destructive file actions, missing checkpoint artifacts, and unclear autonomy boundaries.

See docs/audit-offer.md.

Install From Source

python3 -m venv .venv
. .venv/bin/activate
python -m pip install -e .
askit --help

Test

python -m unittest discover -s tests

Safety Defaults

• local files only unless the user wires a separate integration
• no external sends by default
• no secret values in reports
• no account, auth, or permission changes
• generated reports are ignored by git by default

Launch Status

This repository is in early public-core preparation. The public CLI is intended to stay small and auditable. Pro modules should live in a separate private repository.

License

The public core is Apache-2.0. Pro modules, templates, and private adapters are distributed under a separate commercial license.