Stop AI-generated code from shipping security and architecture problems
Norma gives Cursor, Claude Code, Lovable and Replit your rules before they write, then checks every change before it ships. Free, permanently.
Which one are you?
For builders
Is my app safe to ship? Scan a repository free and get the fix prompt.
Go to the builder page →
For engineering teams
Norma checks every pull request your agents open and comments with what breaks your rules.
Go to the engineering page →
For CTOs and the enterprise
Rules, approvals and an audit record for every AI change.
Go to the enterprise page →
Your agents write faster than you can review
Teams using coding agents get more pull requests than anyone has time to read. Norma checks each one against your rules, and the GitHub App comments with every new issue it finds. Reviewers look at what failed instead of reading every line, and the whole team sees the same result.
See it working before you connect anything
Three ways to watch Norma run without signing up and without pointing it at your own code.
Interactive demo
A guided walkthrough of the loop. The agent proposes, Norma returns the rules that apply and the violations, the agent corrects, and the change lands. Runs in your browser, no install.
Open the interactive demo →
Live demo repository
A public repository checked by Norma. Read the code, the checks it runs and the findings history, all in the open.
Open the demo repository →
The demo and the repository are read only. Nothing reaches your own repositories until you connect one yourself.
Norma MCP
Ask your agent if it’s ready to ship. Get the same answer every time.
Connect Norma to Claude Code, Cursor, Copilot or any MCP client. Your code gets checked against your rules while the agent writes, and the result comes back the same whoever asks and whenever they ask.
Any coding agent can tell you what a function does, and a rules file can tell it what your standards are. The verdict is the harder part. A model asked twice about the same file can answer twice differently.
Norma runs the check outside the model and returns a fixed result: rule ID, severity, line, fix.
A verdict that repeats. Same file, same rules, same answer, whichever agent is driving.
A record you did not write yourself. Rules verified, violations fixed and where, tagged with the model and version.
One ruleset, versioned centrally, so 40 repositories cannot drift into 40 versions of your standards.
Under 30 seconds. OAuth sign-in, one config entry.
norma · mcp session
Is admin-client ready to ship?
norma · live_check → src/lib/admin-client.ts
No. Two rules fail, 10 pass.
HIGH
line 15
sb-exposed-service-role-key-1.0
A service_role key is hardcoded in client code. It bypasses every row level security policy. Move it into an Edge Function.
MEDIUM
line 15
sb-hardcoded-anon-key-1.0
The key is pinned in source, so rotating it means a code change. Read it from an environment variable.
Want me to fix both?
Yes
norma · live_check → clean
norma · register_applied_actions → 2 fixed, 10 verified
Both fixed and re-checked. Recorded against your ruleset, tagged to the model that made the change.
Same file, same rules, same verdict, on any agent.
Before the agent writes
Your agent asks Norma for your rules before each suggestion. The agent writes, Norma checks, and anything that fails goes straight back to the agent. What comes out is working code plus a record of what was checked and fixed. Unlimited checks on the free tier, no usage cap.
After the code lands
Connect GitHub or Bitbucket and run a deterministic scan of the whole repository. Norma discovers your stack, frameworks and integrations automatically. Weekly Full Scans, processed in memory.
One rule library, and the share of your code a model wrote
One rule library
Rules grounded in SOLID principles and stack best practice. You switch off what does not fit your context. The same set runs while the agent writes and in every Full Scan, and every finding records rule, version, file, line, commit, author and timestamp.
AI Authorship Ratio
Know what share of your codebase was written by a model, and which band you are in.
The free tier is permanent, not a trial. Pro at $199 per month, and Business is quoted for your organization.
What is an AI coding agent guardrails tool?
It is software that sets the rules an AI coding agent has to follow and checks the agent’s output against them before the code lands. Norma is one. It gives Claude Code, Cursor or any MCP capable agent your organization’s rules before it writes, runs the check outside the model so the same file always gets the same verdict, and records what was checked and what was fixed. A rules file asks the agent to comply. Norma checks whether it did.
What is AI code governance?
It is the practice of deciding what good code looks like for your organization, then enforcing that decision automatically wherever code gets written, including when a model writes it. Rules are authored in plain English, reviewed and approved, versioned so every change is traceable, and applied identically in every editor and every scan. What you get back is not just a warning, it is a record of what was enforced and why.
Why does the same code always get the same result?
Norma runs the check outside the model, so the same file against the same rules returns the same rule ID, severity, line and fix every time. A judgment that changes between runs cannot be a gate: if a file passes on Tuesday and fails on Wednesday with nothing changed, the team stops trusting the result. A fixed result can be trended across releases, used as a Quality Gate and handed to an auditor.
Which editors and agents does Norma work with?
Cursor, Claude Code, Windsurf, VS Code, Lovable, Replit and any other MCP capable client. Add one entry to your MCP config and the first call opens your browser to sign in. OAuth, no API key needed. Agents your own team builds can call the same server.
Who decides which rules apply?
You do. Norma ships a rule library grounded in SOLID principles and stack best practice, and your architects decide what applies: switch off the rules that do not fit your context, and every check runs against the set you chose. What ran and what was excepted is always on the record.