Connect
The server’s address is https://nestedbit.com/mcp. It uses Streamable HTTP and needs no sign-in or key.
Claude Code
claude mcp add --transport http nestedbit https://nestedbit.com/mcp
Cursor, in .cursor/mcp.json or ~/.cursor/mcp.json
{
"mcpServers": {
"nestedbit": {
"url": "https://nestedbit.com/mcp"
}
}
}
VS Code: run MCP: Add Server, choose HTTP, and enter the address.
Other clients: add a remote MCP server with the address above. It serves MCP 2026-07-28 and older versions.
Tools
| Tool | What it does |
|---|---|
check_mcp_server |
Checks a public MCP server’s sign-in chain, how it treats fake tokens, and what an open server exposes. |
get_mcp_fix |
Returns code for a finding, for the TypeScript SDK, the Python SDK or FastMCP. |
review_mcp_tools |
Sorts the output of tools/list by risk, and flags hidden instructions and misleading annotations. |
inspect_jwt_claims |
Checks a JWT’s decoded header and payload. It doesn’t take the token or its signature. |
export_rbac_policy |
Turns a role model or template into Casbin, OPA Rego, Cedar, Postgres RLS, CSV, Markdown or JSON, and checks it. |
lint_supabase_rls |
Checks a Supabase schema, as SQL or the catalog query’s JSON, for tables and functions the Data API exposes, and returns SQL that fixes each finding. |
generate_supabase_rls |
Writes row-level security policies, or SQL tests for them, from who may do what on each table. |
lint_firestore_rules |
Checks Firestore or Cloud Storage rules for open access, test-mode dates and missing validation, path by path. |
generate_firestore_rules |
Writes Firestore rules, emulator tests or a firebase.json from who may do what in each collection. |
explain_k8s_rbac |
Flags risky Kubernetes Roles and bindings, answers “who can do this?”, and says what one user, group or service account may do. |
compare_auth_providers |
Estimates the monthly cost of 13 auth providers at your numbers, and which of your needs each one meets. |
score_ai_readiness |
Scores an organization across six areas and names what to fix first. |
rank_ai_workflows |
Ranks repeated work by how well it suits an AI agent, and suggests the kind of agent. |
Every tool only reads what it’s given and returns text. None of them changes anything, and
check_mcp_server never calls a tool on the server it checks. The same tools are on the
free tools page if you’d rather use them in a browser.
What happens to what you send
On the website, most of these tools run in your browser. Here, what your client sends goes to our server on Google Cloud Run in the United States. We use it to answer and don’t save or log it. We count how often each tool is called, by its name alone. The privacy page has the details.
Don’t send passwords, API keys, signed tokens or real personal data. The schema, rules and RBAC tools need settings
only: send a schema without its data, and RBAC objects rather than Secrets. Each paste can be up to about 250 KB.
For a JWT, have your client decode it first:
inspect_jwt_claims refuses anything that looks like a signed token. To test a token’s signature or look
for a weak secret, use the JWT Security Inspector, which runs in your
browser and sends nothing.
Safe to call from an agent
A checked server could fill its metadata and tool descriptions with instructions aimed at your model, so the results leave that text out. You get our findings and fixes, and any tool name that looks unusual is replaced with a placeholder. The report on the website shows the full details to you, not your model.
The same goes for what you paste. A schema, rules file or manifest often comes from a repository or a cluster someone
else wrote, and its names are quoted throughout the results. So a name that doesn’t look like a name, with quotes,
odd characters or words aimed at a model, comes back as unusual_name_1 and so on, and parser errors leave
out the text they couldn’t read. Ordinary names, including policy names like “Enable read access for all users”,
come back as written.
Requests from web pages on other sites are refused. Calls are limited to 60 a minute per client, and server checks to 5 a minute and 30 an hour, with 6 a minute for any one server. Each client also gets 50 checks and 500 other tool calls a day, reset at midnight UTC, with a daily total across all clients. Everyone behind one address, such as an office or a gateway, shares these limits. Contact us if you need more.
What our own checker says about it
Run the MCP Security Checker on this server and it reports a medium finding: check_mcp_server takes a
free-form URL. That’s accurate, and it’s why the tool’s server refuses private and internal addresses, but that
protection can’t be seen from outside, so the finding stays. It also notes that the server is open, which suits it:
it holds no private data and needs no sign-in.
Need this for your own server?
We build and audit MCP servers, including OAuth sign-in and per-tool permissions. Book an intro call