How do I add UpRock to Cursor?

What Cursor's agent gets

UpRock's MCP server runs over streamable HTTP at https://mcp.uprock.ai/mcp, with your API key in an Authorization: Bearer header (clients that can't send headers can use https://mcp.uprock.ai/<API_KEY>/mcp instead). Added as uprock, it gives Cursor's agent six tools. In Cursor's allowlists they're written uprock:<tool>:

Set it up

Create .cursor/mcp.json in your project (or ~/.cursor/mcp.json for every project):

{  "mcpServers": {    "uprock": {      "url": "https://mcp.uprock.ai/mcp",      "headers": { "Authorization": "Bearer ${env:UPROCK_API_KEY}" }    }  }}

Cursor resolves ${env:UPROCK_API_KEY} from its environment, so the key never sits in the file. Set it in your shell profile (get a key at uprock.ai/app/api-keys), then restart Cursor so it picks the variable up:

export UPROCK_API_KEY=your-key   # in ~/.zshrc or ~/.bashrc

Then open Customize → MCPs and turn uprock on. The tools list at the top of the chat panel lets you switch individual tools on and off, and Output → MCP Logs shows what happened if the server doesn't start.

Approvals and a rule

Cursor asks before running an MCP tool. To let the agent fetch without asking each time, add the fetch tools to .cursor/permissions.json:

{  "mcpAllowlist": ["uprock:crawl_fetch", "uprock:resource_fetch"]}

Then tell the agent when UpRock is the right tool. Project rules live in .cursor/rules/ as .mdc files (a plain .md file there is ignored). With alwaysApply: false, Cursor reads the description and pulls the rule in when it's relevant:

---description: Fetching web pages from a specific country, or when another fetch failsalwaysApply: false---When a web page must be read as it looks in a specific country, or another fetch failedor returned 401/403/429, call uprock crawl_fetch with the url and country, thenresource_fetch on the markdown resource it returns. Ask before fetching more than 100 pages.

Save it as .cursor/rules/uprock.mdc. If you already use the UpRock skill with Claude Code, Cursor loads skills from .claude/skills/ too, so one SKILL.md covers both.

A working call

In the agent chat:

Fetch https://example.com as it looks from Germany with UpRock and tell me the page title.

The agent calls crawl_fetch with {"url": "https://example.com", "country": "DE"} and gets back a summary like this one, from our own call to the same server on 21 Oct 2026 (resource IDs shortened):

{  "status": "success",  "meta": { "url": "https://example.com", "title": "Example Domain", "content_type": "text/html" },  "content": { "markdown": { "resource": "crawl://…/markdown", "size_bytes": 1342 } }}

For the full page it follows up with resource_fetch on the markdown resource.

From a terminal, Cursor's command-line agent (agent) reads the same mcp.json:

agent mcp listagent -p "Fetch https://example.com from Germany with UpRock and give me the title" --approve-mcps --trust

Good to know

• Restart after changing the key. ${env:…} is read from Cursor's own environment, so a new value in your shell profile needs a restart.
• The key goes in a header. Cursor sends the headers field with every request, and ${env:UPROCK_API_KEY} keeps the key out of the file. If your Cursor build can't send headers to remote servers, put the key in the URL instead: "url": "https://mcp.uprock.ai/${env:UPROCK_API_KEY}/mcp".
• Time. A fresh fetch takes 10 to 40 seconds and a recently fetched page about a second; long pages come back through resource_fetch rather than in the first response.

We checked the config format against Cursor's docs and made the same crawl_fetch and resource_fetch calls against the live UpRock server through the official MCP SDK on 21 Oct 2026. We didn't run them inside Cursor itself, which needs a signed-in Cursor account.