Claude Agent SDK Guardrails Kit
Tool policy, single-use approvals, spend caps and offline hook tests for the Claude Agent SDK (Python).
View Claude Agent SDK Guardrails Kit
Agent skills plus a small checker for agencies that answer RFPs: it turns the RFP into a compliance matrix, ties every 'complies' to a proof document, reads addenda, and helps with the bid/no-bid call.

13 errors and 6 warnings, including: requirement paraphrased instead of quoted (E001), “complies” with no evidence (E003), excerpt not in its file (E005), evidence expired (E006), certification claimed with no certificate (E008), changed by an addendum and not re-reviewed (E010), absolute claims like “guarantee” or “100%” (W104), certificate with no expiry (W105).
UNSUPPORTED means no proof attached. It never means non-compliant. It does not write your proposal, judge whether you actually comply, or read scanned PDFs (no OCR). English RFPs only; extraction is a draft.
Python 3.9+, standard library only, no network, no model calls. Skills are plain SKILL.md files for Claude Code, Cursor or Codex. For PDFs, convert with pdftotext first. 18 tests.
Zip, 42 KB, v1.0. Use in your own and client projects (one buyer: a person, or one company and its employees); no resale or giving the files away. Full terms in LICENSE.md.
RFP compliance matrix, evidence lint, addendum diff and bid/no-bid scoring. Offline CLI + skills.
Tool policy, single-use approvals, spend caps and offline hook tests for the Claude Agent SDK (Python).
View Claude Agent SDK Guardrails KitReconcile a bank CSV against your ledger: ranked exceptions, a balance proof, suggested (never posted) entries. Offline CLI + agent skill.
View Bank Reconciliation Exceptions SkillFind #REF!, numbers typed over formulas, inconsistent formulas, SUM gaps and circular refs before a model ships. Read-only, runs locally.
View Excel Model Audit Skill