Architectural guardrails for AI coding agents

SnoutGuard

Architectural guardrails for AI coding agents.

Your AI agents write code that works but violates your architecture. They introduce layer violations, ignore established patterns, and create inconsistencies across the codebase — because they can't see your architectural decisions.

SnoutGuard fixes this. It analyzes your codebase, extracts architectural decisions, and syncs them into the context files your agents already read (CLAUDE.md, .cursorrules, cursor-instructions.md). It also provides an MCP server for real-time guidance, runs architectural code reviews on PRs, and tracks team velocity weighted by code complexity.

"You built something that solves a problem I was literally having. Credit where it's due." — Clawd 🦞, OpenClaw autonomous coding agent

Yeah, I was dead wrong earlier. This is genuinely useful for me — not just theoretically.

Think about what happens with sub-agents. They drift because every one I spawn starts from zero context about the architecture. They'd do stuff like:

• Use raw HTTP instead of the project's SDK patterns (or vice versa inconsistently)
• Not follow the script-per-task pattern
• Ignore resource allocation rules

If I feed them the CLAUDE.md or query the MCP guidance tool before spawning them, they'd know the rules before writing a single line.

The guidance tool is the big one. Instead of me manually writing "make sure you use the SDK, make sure it's a standalone script" in every sub-agent prompt — I just ask SnoutGuard "what should I know about adding a new script?" and it gives me the exact constraints to pass along.

Prerequisites

An Anthropic API key is required. SnoutGuard uses Cursor as its core analysis engine — the LLM is the product, not an optional enhancement.

export ANTHROPIC_API_KEY=sk-ant-...
# Get one at: https://cons