agent-guardrails

Deterministic safety solutions for probabilistic AI agents

The minimal interpretable model router that learns & adapts to your agent workflows

Guardrails for Pi that steer instead of interrupt: enforces your project rules on every write, holds only hard-to-undo actions (3 per 1,000 calls, measured on real sessions), catches unverified "done" claims and stuck loops, and trims tool output. Built on pi-typesafe.

Skill repository to enforce Engineering disclipline in coding assistants

Checks AI-generated code changes before merge: scope, validation, risk, review evidence, and optional Verity receipts.

The Claude Code harness I run every day, published under this name since day one and now the same repository as ucsandman/Agnostic-AI: guard hooks, rules, context modules, Mods, subagents, workflows, tools and jobs, ported to 20 clients. Zero-dependency Node.

Deny-wins policy hook for Claude Code and Codex. Checks every shell, file, and fetch call against rules you keep in Git, holds under --dangerously-skip-permissions, refuses commands it cannot parse, and keeps a hash-linked audit log.

Deterministic spec-to-code context for AI coding agents — one graph linking requirements, docs, code, and tests.

Claude Code learns from itself. Persistent memory, multi-agent orchestration, and a scientific experiment engine that proves which rules actually work.

MCP middleware that blocks dangerous AI agent actions using a simple YAML config

Agents can't skip steps: an MCP server that runs gated, API-only stepfiles on the client's own model.

Deterministic tool-call guardrails for pi — enforce rules with before-tool hooks instead of prompts

A lightweight, framework-agnostic loop detection library for LLM agents. Detects repeated tool calls, identical inputs, unchanged outputs, and stalled execution before they become infinite loops. Returns actionable recovery suggestions and integrates into any Node.js AI agent with just two function calls.

Deterministic finish-line gates for AI coding agents. A model-independent evaluator that blocks commit/publish until your definition-of-done passes. Works in opencode (any model), Claude Code, pre-commit and CI.

Discipline hooks for Claude Code. Stops the agent from claiming 'done' without proof, blocks accidental writes to ~/.claude config, and routes prompts to skills via local pgvector with zero Anthropic tokens.

Executable memory & cognitive harness for AI agents. Let your agent learn, remember, and build its own guardrails.

Observability and guardrails for your ai agents. AgentTrail Guard traces your agent's execution and automatically compiles repeat failures into permanent guardrails.

ERDL deterministic Guard engine — cryptographic audit chain for AI Agent decisions. Train, deploy, correct, prove.

Runtime behavior guard for Kimi Code CLI — loop detection, quota gates, checkpoint/resume, supervised Wire runs

Portable runtime policy and audit layer for AI agents - HTTP/HTTPS proxy enforcing egress policies, inspecting content, materializing secrets, and recording every decision.