Radar MCP server
The Radar MCP server lets you ask an AI client, such as Claude, about the tool lists of the official vendor MCP servers that Radar follows: “did Hugging Face’s MCP server change its tools this week?”, or “which of Firestore’s tools can delete things?”. It answers from the same published data as the tracker pages.
It is read-only and needs no sign-in. It cannot act on any service, it does not connect to a vendor’s server for you, and it does not say whether a server is safe. Answers give tool names as identifiers and never quote a server’s descriptions.
Radar is separate from the Stackduke gateway that the rest of these docs describe. That gateway, which will check agents’ tool calls, is an upcoming product planned for the invite-only beta in March 2027, and it is not part of this server. Nothing Radar answers is a Stackduke decision about a tool call.
The address
Section titled “The address”https://radar.stackduke.com/mcpIt speaks MCP over Streamable HTTP. There is no account, token, header or OAuth step.
Add it to Claude
Section titled “Add it to Claude”-
Add a custom connector. In Claude, open Settings, then Connectors, and choose Add custom connector. On a Team or Enterprise plan, an owner adds it for the organization from the organization’s connector settings.
-
Enter the address. Name it
Stackduke Radarand enterhttps://radar.stackduke.com/mcp. Leave the advanced settings (OAuth client ID and secret) empty: Radar has no sign-in. -
Turn it on in a chat. In a chat, turn on Stackduke Radar from the tools menu, then ask a question such as “Which MCP servers does Stackduke’s tracker follow?”.
Expected result: Claude calls list_tracked_servers and answers with the tracked servers, each with its tool count and when it was last checked, and a link to the tracker page. Your client may ask you to allow each tool the first time it is used. All four only read tracker data.
If it does not connect: check the address ends in /mcp, and that no OAuth details were entered. The address https://mcp.stackduke.com is a different service, the planned gateway, and is not Radar.
Add it to another MCP client
Section titled “Add it to another MCP client”Any MCP client that supports remote servers over Streamable HTTP can use Radar. Give it the address above, with no authentication. For example, in Claude Code:
claude mcp add --transport http stackduke-radar https://radar.stackduke.com/mcpFor other clients, add a remote (HTTP) server in the client’s MCP settings, with the address above. The settings format differs between clients, so follow your client’s own instructions.
The four tools
Section titled “The four tools”| Tool | Asks for | Answers with |
|---|---|---|
list_tracked_servers |
nothing | The tracked servers, each with its tool count, its tools counted by kind, when it was last checked and when its tools were last listed |
get_server_capabilities |
a vendor, and optionally a kind | The tool names on that vendor’s server, grouped by the most serious thing each is counted as able to do |
get_recent_changes |
a vendor, and optionally a number of days from 1 to 30 (default 7) | The changes seen in that period: tools added and removed, tools changed with the fields that changed, and tools now counted as able to do more. Also how many checks were made, which failed, and any stretch that was not watched |
find_tool |
a tool name, and optionally contains to match part of a name (at least 3 characters) |
Which tracked servers list a tool with that name, and the kind each is counted as |
A vendor can be given by id or name, for example google-firestore or Hugging Face. Every answer ends with a link to the tracker page it comes from, and says when the data was published.
The kinds
Section titled “The kinds”The kinds come from Stackduke’s open-source MCP scanner. Each tool is counted once, under the most serious thing it is counted as able to do, in this order:
| Kind | Meaning |
|---|---|
money |
can move or spend money |
destructive |
can make changes that cannot be undone, such as deleting |
send |
can send messages or publish to other people |
write |
can create or change things |
unclear |
could not be placed from its name or annotations |
read |
labelled as only reading data |
The labels come from the server’s own annotations where it sends them, and otherwise from a guess based on the tool’s name that leans to the more serious kind. They are not verified by calling the tools. They are the scanner’s reading, not a safety rating.
Limits
Section titled “Limits”- Only the servers Radar follows. These are the official MCP servers on the tracker, each of which lists its tools with no sign-in. Radar reads each one the way any MCP client would, with no account. Any other server gets the same “not covered” answer, which says nothing about that server either way.
- About once a day. The tracker data is published about once a day. A change is dated as seen between two checks, and a change made and undone between two checks is not seen. If the data is more than 36 hours old, every answer says so.
- Recent changes only. The data keeps the five most recent changes for each server. The tracker page for each server has the full history.
- No vendor text. Radar never quotes a tool’s description, title or schema. It can say which fields changed, not what they say now.
- Not a safety rating. Radar never says a server is safe, or what a server cannot do.
- Requests. Up to 300 requests a minute from one address. Past that, the answer is HTTP 429, and the client can try again after a minute. AI clients often call from shared addresses, so the limit is set high.
Privacy
Section titled “Privacy”- What reaches Radar. Only the tool calls your client makes: the tool name and its arguments, such as a vendor name or a number of days. The rest of your conversation is not sent to Radar; it stays between you and your AI provider, under the provider’s terms.
- Not logged or kept. Radar logs nothing about requests: not the arguments, not which tools were called, not client details. It stores nothing, and sets no session or cookie.
- Your IP address. Cloudflare’s rate limiter counts requests by the calling IP address. Radar uses it for that count only and does not keep or log it.
- Cloudflare. Radar runs on Cloudflare Workers. Cloudflare processes request metadata to deliver and protect the service, and keeps aggregate request and error counts.
See the Privacy Policy for the full details.