Destructive shell access
Broad Bash rules, force pushes, recursive deletion, hard resets, and remote scripts piped to a shell.
Free private scanner
Find the permissions your coding agent should never have.
Scan agent instructions and MCP configuration for destructive commands, broad permissions, exposed secrets, and missing approval gates. Everything runs locally in your browser.
No uploadNo accountActionable findingsFour tool formats
Scan your agent configuration
Best results: add every repository instruction, settings, rules, and MCP file together.
Files stay on this device
{ }
Your findings will appear hereChoose files or paste a configuration. Nothing is uploaded.
What it checks
Agent instructions are part policy, part executable attack surface.
The scanner looks for high-risk permission patterns and for the controls that are often missing when a coding agent can use the terminal, network, credentials, and third-party tools.
Secret exposure
Likely live tokens, private keys, and hard-coded MCP environment credentials.
MCP supply-chain risk
Plain HTTP endpoints, mutable package execution, and unpinned npx or uvx servers.
Missing approval gates
No explicit pause before deploys, data deletion, billing changes, messages, or other external effects.
Unbounded scope
No repository boundary, protected-path rule, or instruction to preserve unrelated work.
Unsupported success claims
No requirement to verify tests, deployments, payments, or external changes with direct evidence.
Automated fix kit
Turn the findings into controls your repository can enforce.
The AI Coding Agent Safety Kit adds conservative cross-tool instructions, a real CI check, a review gate for protected paths, and the decision records teams usually leave implicit.
$39 one payment · instant ZIP
Get the complete safety kitSecure Stripe checkout · instant download · email backup · no subscription
AGENTS.mdCross-tool safety boundariesNegative constraints, scope rules, approval gates, and verification.CLAUDE.mdClaude Code controlsTool-use boundaries and untrusted-instruction handling..cursor/rulesAlways-on Cursor ruleMatching safety policy in Cursor's native format.GitHub ActionsDeterministic CI enforcementSecret patterns, unsafe config, and protected-path review.PR templateAgent disclosure and evidenceReview checklist with explicit verification fields.MCP checklistServer approval recordProvenance, capabilities, data handling, and isolation.Permission matrixEnvironment-specific limitsConcrete local, CI, staging, and production decisions.Incident runbookContainment and recoveryA concise response sequence for unauthorized agent actions.Questions
Before you scan
Do my files leave the browser?
No. The scanner uses browser JavaScript to read the files you select. It has no upload step, sends no file contents to BurnRate, and stores no scan data.
Which tools and files are supported?
The rules cover AGENTS.md, CLAUDE.md, Cursor `.mdc` rules, Claude settings, MCP JSON, and similar Markdown, JSON, YAML, TOML, or text configurations. Unknown text files still receive the generic secret and dangerous-command checks.
Does a low score mean my setup is secure?
No. This is a focused heuristic check, not a penetration test, compliance assessment, or guarantee. It cannot inspect runtime behavior, organization settings, or files you do not select.
Will the paid kit work without a subscription?
Yes. It is a one-time ZIP download. The shell check and GitHub workflow run in your repository without a BurnRate account or hosted service.
Can I share the kit across my company?
The license covers one individual or one organization and repositories it owns or controls. It does not permit resale, public redistribution, or sublicensing.
This tool and kit provide general technical starting points, not legal, security, compliance, or professional advice. Review and test controls for your own environment.