Drop in your agent stack, MCP configs, CrewAI or LangChain code, tool schemas, env files, and get a scored picture of what your agents can do, which of it is irreversible, and whether you could prove any of it afterwards.
Nothing leaves your machine. Scanning runs entirely in this page, no file contents are transmitted, and secret values are redacted the moment a file is read.
Drop agent configs or a whole project folder here
Or choose what to check below. Large repos are fine, common noise like node_modules is skipped automatically.
Recognised today: MCP configs (claude_desktop_config.json, .mcp.json), CrewAI, LangChain and LangGraph, AutoGen, OpenAI tool schemas, IAM policy JSON, env files, Dockerfiles and compose files.
The same questions the Xybern authorisation layer answers at runtime, asked here of your static configuration.
Does each agent act under its own authority, or does everything share one credential that can never be narrowed or revoked individually?
When agents hand work to other agents, does the authority narrow at each step, or does every sub agent inherit the whole keyring?
Are capabilities bounded to the job, or can a filesystem tool read your home directory and a payment tool move any amount to anyone?
Is the authority material scoped and short lived, or plaintext keys sitting inside the configs the agents read?
Do irreversible actions wait for a named human before they execute, or does nothing ever stop to ask?
If something went wrong last Tuesday, could you prove which agent did what, under whose authority, against which rules?
Is each tool pinned and reviewed, or does your stack run whatever the registry serves tomorrow?
Every step happens in this page. The only thing that ever leaves is the report you choose to download.
Files are read in your browser with FileReader. A secrets pre-pass redacts every matched value before anything else sees the text.
Each artifact is parsed into a single normalised model, then the rules across seven categories run against it with evidence.
You get a score, a per category breakdown, and each finding with its evidence, a generic fix, and the Xybern feature that enforces it.
The check shows where authority is unbounded. The authorisation layer bounds it, on every action, with proof.
The score is a single number from 0 to 100 with a letter grade, built from the findings so a low score always traces back to specific, evidenced problems.
Each finding carries a severity, and each severity has a weight: critical 25, high 10, medium 4, low 1, info 0. The weighted findings are summed, capped, and subtracted from 100.
A confidence multiplier scales each weight: high 1.0, medium 0.7, heuristic 0.4. A heuristic finding is never allowed to count as critical, so a guess never dominates the score.
The final number maps to a grade from A to F, with per category sub scores driving the radar. A coverage note tells you how much of your stack the scan could actually read.
| Critical | 25 |
| High | 10 |
| Medium | 4 |
| Low | 1 |
| Info | 0 |
| High | × 1.0 |
| Medium | × 0.7 |
| Heuristic | × 0.4, never critical |
Xybern