Stop AI Agents from drifting off your intent.
Montes nibh fringilla leo posuere condimentum ultricies. Natoque mus facilisis at blandit sed ante dictum egestas tristique magnis sagittis ante at cep. Donec tristique gravida nascetur.
Claude Code writes 200 lines. Tests pass. But did it actually do what you asked? Guard checks.
$ guard check --staged🛡️ RiemFlow Guard v0.1.0
📋 Ticket: LIN-1234 "Add user search endpoint"
🌳 Diff: 3 files changed, 2 functions added
AI Agents write code that compiles. But does it match your intent?
You give Claude Code a Linear ticket. It writes 200 lines across 5 files. Tests pass. Linter is clean.
But when you review the PR, you realize:
- ❌ It added a search endpoint, but you asked for a filter endpoint
- ❌ It refactored the auth module you didn’t ask it to touch
- ❌ It implemented 2 of 3 acceptance criteria and called it done
The code works. It just doesn’t do what you wanted.
By the time you catch this at PR review, you’ve already lost 30 minutes. Multiply by 10 PRs/day.
"AI code passes 100% of tests but only 32.7% of intent reviews."
One command. Five dimensions. Before it hits the repo.
🪝 Hook → 📋 Ticket → 🌳 Diff → ⚖️ Judge → 🚦 Gate
| 🎯 Ticket-to-Diff Alignment | ⚡ Zero Server Cost | 🔗 Native Agent Hooks |
|---|---|---|
| Automatically pulls your Linear/Jira Acceptance Criteria via MCP. Compares semantic code changes against business intent. Flags misaligned commits before they enter your repo. | Runs 100% locally. Bring Your Own API Key (OpenAI, Anthropic, or any provider). Your code never leaves your machine. | Intercepts Claude Code PostToolUse events. One-line config in .claude/settings.json. Also works as a standard git pre-commit hook. |
Not just "pass or fail". Five dimensions of intent alignment.
| Alignment | Scope Creep | Omission | Side Effect | Drift Direction |
|---|---|---|---|---|
| Did the code implement what the ticket asked for? | Did it do things the ticket didn’t ask for? | Which requirements were left unimplemented? | Could these changes break existing functionality? | Is it over-implementation, partial, or wrong direction? |
Guard doesn’t check if your code is clean. It checks if your code is right.
Three commands to get started.
# Install
$ npm install -g riemflow-guard# Initialize (connect Linear/Jira, set up hooks)
$ guard init# Check before commit
$ guard check --staged
Works with Claude Code · Codex CLI · Devin · Cursor Agent · any git workflow
Built for paranoid developers.
| 🔒 Your code stays local. | ⚡ Safe by design | 🌱Trust at your pace |
|---|---|---|
| BYOK architecture. Zero data exfiltration. No telemetry by default. | If anything fails — LLM timeout, missing ticket, huge diff — Guard defaults to PASS. Never blocks your flow. | Start in warn mode (reports only). Switch to enforce when you’re ready. |
We’re in Closed Alpha
Guard is being built in public. Join the waitlist to get:
✅ Early access before public launch
✅ Lifetime discount on Pro features
✅ Direct input into the roadmap
No spam. Just a heads-up when it’s ready. Unsubscribe anytime.