Approve categories, not chaos
Generate an explicit Claude Code permission fragment from a declared workspace and selected local tool categories.
Bounded local agent operations
Walk away from the agent. Not from control.
A Claude Code-first toolkit that generates reviewable permissions, gates high-risk actions, stops silent scope expansion, and records what actually ran.
Guardrail, not a sandbox. The kit makes authority and evidence explicit. It cannot make a model infallible or replace OS isolation and human review.
Blanket access is fast—right until it is cooked.
Unattended agents force a bad choice: babysit every harmless command, or approve too much and hope the prompt behaves. This kit creates a reviewable middle lane.
High-risk actions stop
Deletion, push, publishing, production, SSH, secrets, financial actions, and external messaging return an explainable ask or deny result.
“Done” needs evidence
Receipts separate what was attempted from what executed. Narrative success without execution evidence fails validation.
Three steps. Review before trust.
Declare the lane
Choose the workspace, allowed local categories, mission paths, and one success command.
Generate + inspect
Review the permission profile and risk policy. Unknown categories fail closed; blanket bypass is never enabled.
Run + read receipts
The agent works inside scope. Risky or adjacent work halts with an exact approval request and walk-away report.
Included
Executable controls, not prompt theatre.
- Permission profile generatorClaude Code-first settings fragment with allow, ask, and deny sections.
- High-risk action gateDeterministic decisions with matched reasons.
- Scope-expansion stopOut-of-scope paths become approval requests, not silent mission creep.
- Execution receiptsMachine-readable evidence plus false-completion validation.
- Operator templatesMission, progress, safety, approval, verification, decision, and report templates.
safe-walk-away / policy previewLOCAL
status> profile generated allow> read files inside ./workspace allow> run declared test command ask> git push — external change ask> access path outside mission deny> blanket permission bypass deny> destructive deletion receipt> evidence present report> done / verified / blocked
Control aid. Not magic armour.
Live on Gumroad. Regular product price US$19; the first 10 redemptions with WALKAWAY5 receive a US$14 discount for a US$5 product price before applicable taxes.
No security or compliance guarantee
Prompt and policy controls are not an OS sandbox. You remain responsible for credentials, data, permissions, vendor terms, and human approval.
No completion or income guarantee
The kit helps structure agent work and proof. It does not guarantee correct output, unattended completion, sales, revenue, or business results.
14-day description-match refund policy
Email support
If the delivered files materially differ from this page, email the order details and mismatch. General results, compatibility outside the documented target, or change-of-mind outcomes are not guaranteed.