How do I give Claude Code live web access with UpRock?

What Claude Code gets

UpRock's MCP server runs over streamable HTTP at https://mcp.uprock.ai/mcp, with your API key in an Authorization: Bearer header. Once it's added as uprock, Claude Code sees these tools, named mcp__uprock__<tool>:

Set it up

Option 1: a project .mcp.json, with the key in your environment (recommended)

Put this in .mcp.json at the root of your project. It's safe to commit: the key stays in your environment.

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

Then set the key in your shell profile (get one at uprock.ai/app/api-keys) and open a new terminal:

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

Claude Code fills in ${UPROCK_API_KEY} on every request. If the variable isn't set, claude mcp list shows a config warning and the server still connects, but every tool call comes back with "An UpRock API key is required".

Option 2: one command, for just you

claude mcp add --transport http uprock https://mcp.uprock.ai/mcp --header "Authorization: Bearer $UPROCK_API_KEY"

This adds the server for the current project only (the default "local" scope, stored in ~/.claude.json), with your key stored as text in that file. Add --scope user to make it available in every project. claude mcp list doesn't show the key, but claude mcp get uprock prints it.

Option 3: the UpRock plugin

The UpRock plugin bundles the MCP server with a few slash commands. Add its marketplace and install it:

claude plugin marketplace add uprockcom/uprock-claude-pluginclaude plugin install uprock@uprock

Inside a session, /plugin marketplace add and /plugin install do the same. Claude Code then asks for your UpRock API key and keeps it in your OS keychain; /uprock:setup checks the connection. If you allow tools in .claude/settings.json as shown below, check the plugin's tool names in /mcp first.

Option 4: add the skill

The UpRock skill teaches Claude when to reach for UpRock and what it costs. Download it from uprock.ai/skill/uprock-skill.zip and unzip it into ~/.claude/skills/, so you end up with ~/.claude/skills/uprock/SKILL.md. Use it together with option 1 or 2.

Check it works

claude mcp list

With option 1, the first run in a project shows the server waiting for approval:

uprock: https://mcp.uprock.ai/mcp (HTTP) - ⏸ Pending approval (run `claude` to approve)

Start claude in that folder and approve the server. This is what the list showed after that on 29 Oct 2026:

uprock: https://mcp.uprock.ai/mcp (HTTP) - ✔ Connected

Inside a session, /mcp lists the server and its tool count.

Tell Claude when to use it

Claude chooses tools by reading their descriptions, and a line in CLAUDE.md makes the choice predictable:

Web pages: use your normal fetch first. If it fails or returns 401/403/429, or I ask for a pageas it looks in a specific country, use UpRock's crawl_fetch with that country, thenresource_fetch on the markdown resource for the full text. Ask before fetching more than 100 pages.

To stop Claude asking for permission on every call, allow the two fetch tools in .claude/settings.json:

{  "permissions": {    "allow": ["mcp__uprock__crawl_fetch", "mcp__uprock__resource_fetch"]  }}

A working call

From a terminal, in the project with the .mcp.json above:

claude -p "Use the uprock crawl_fetch tool to load https://example.com from Germany (country DE) and tell me the page title." \  --allowedTools "mcp__uprock__crawl_fetch"

Claude calls crawl_fetch with {"url": "https://example.com", "country": "DE"}. This is the start of what the tool returned when we made that call on 21 Oct 2026 (resource IDs shortened):

{  "status": "success",  "meta": { "url": "https://example.com", "title": "Example Domain", "content_type": "text/html", "time_ms": 1316 },  "content": {    "markdown": { "resource": "crawl://…/markdown", "size_bytes": 1342, "mime_type": "text/markdown" },    "html": { "resource": "crawl://…/html", "size_bytes": 2358, "mime_type": "text/html" }  }}

The summary is small on purpose. When Claude needs the whole page, it calls resource_fetch with the markdown resource.

Limits worth knowing

• Output size. Claude Code warns when a tool returns more than 10,000 tokens and cuts output at 25,000 by default. resource_fetch on a long page can pass that; raise it with MAX_MCP_OUTPUT_TOKENS. Text results over 50,000 characters are saved to a file instead.
• Time. A fresh fetch takes 10 to 40 seconds, and a recently fetched page about a second. Claude Code moves calls that run past two minutes to the background in interactive sessions.
• Where the key ends up. With option 1 it stays in your environment. With option 2 it's stored as text in ~/.claude.json, and claude mcp get uprock prints it, so keep that output out of issues and screenshots. The plugin keeps it in your OS keychain.

On 29 Oct 2026, with Claude Code 2.1.312, both option 1 and option 2 showed ✔ Connected to the live server; pointed at a local test server, option 1 sent the header with the variable filled in. We installed the plugin with the two commands above, and made crawl_fetch calls against the live server with the key in the header. Earlier, on 21 Oct, we made the same crawl_fetch and resource_fetch calls with the key in the URL through the official MCP SDK. Our test environment couldn't run a full Claude session, so the claude -p step follows Claude Code's docs.