How do I give a CrewAI agent web access with UpRock?

What your crew gets

UpRock's MCP server runs over streamable HTTP at https://mcp.uprock.ai/mcp, with your API key in an Authorization: Bearer header. MCPServerAdapter turns each of its tools into a CrewAI tool with the same name, and takes a list of names to load only some of them:

Get a key at uprock.ai/app/api-keys and set it as UPROCK_API_KEY.

MCPServerAdapter or the mcps field

CrewAI's docs recommend the mcps field on Agent for most MCP servers. It works with UpRock as long as the key goes in a header, but MCPServerAdapter is still the better fit. This is what each setup gave us:

A fresh UpRock fetch takes 10 to 40 seconds, so through mcps some will be cut off. The last row is why the key never goes in the URL with CrewAI: CrewAI builds tool names from the URL, path included, so the key would go to your model provider with every request, show up in verbose logs and in CrewAI's "MCP Connected" panel, and the model couldn't tell the tools apart.

The code

The example makes a direct call first, which needs no model, then gives the same two tools to an agent in a crew:

# pip install "crewai[anthropic]" "crewai-tools[mcp]"   env: UPROCK_API_KEY, ANTHROPIC_API_KEYimport astimport jsonimport os​from crewai import LLM, Agent, Crew, Taskfrom crewai_tools import MCPServerAdapter​UPROCK = {    "url": "https://mcp.uprock.ai/mcp",    "transport": "streamable-http",    "headers": {"Authorization": f"Bearer {os.environ['UPROCK_API_KEY']}"},}MODEL = "anthropic/claude-sonnet-5-5"  # any model CrewAI supports​​def summary(output: str) -> dict:    """crawl_fetch returns a JSON summary and two resource links. The adapter keeps the    text and returns str() of a list when there's more than one block."""    text = ast.literal_eval(output)[0] if output.startswith("[") else output    try:        page = json.loads(text)    except ValueError:        page = {}    if page.get("status") != "success":        raise RuntimeError(f"UpRock: {text}")    return page​​with MCPServerAdapter(    UPROCK, "crawl_fetch", "resource_fetch", connect_timeout=60) as tools:    crawl_fetch, resource_fetch = tools["crawl_fetch"], tools["resource_fetch"]​    # 1. A direct call, no model involved: load a page on a device in Germany.    page = summary(crawl_fetch.run(url="https://example.com", country="DE"))    markdown = resource_fetch.run(uri=page["content"]["markdown"]["resource"])    print(page["meta"]["title"], len(markdown), "chars of markdown")​    # 2. The same tools in a crew.    reader = Agent(        role="Web reader",        goal="Read web pages as they look in the country the user names",        backstory=(            "When you need a page from a specific country, or another fetch failed, "            "call crawl_fetch with the url and country, then resource_fetch on the "            "markdown resource it returns."        ),        tools=tools,        llm=LLM(model=MODEL),    )    task = Task(        description="What does https://example.com say when opened from Germany?",        expected_output="The page title and one sentence on what the page says.",        agent=reader,    )    print(Crew(agents=[reader], tasks=[task]).kickoff())

Keep the crew inside the with block: the adapter closes the connection when the block ends.

In a crew made with crewai create

Projects scaffolded with crewai create crew use a @CrewBase class. Give it the server once and pick tools per agent:

import os​from crewai import Agentfrom crewai.project import CrewBase, agent​​@CrewBaseclass ResearchCrew:    mcp_server_params = {        "url": "https://mcp.uprock.ai/mcp",        "transport": "streamable-http",        "headers": {"Authorization": f"Bearer {os.environ['UPROCK_API_KEY']}"},    }    mcp_connect_timeout = 60​    @agent    def reader(self) -> Agent:        return Agent(            config=self.agents_config["reader"],            tools=self.get_mcp_tools("crawl_fetch", "resource_fetch"),        )

Add your @task and @crew methods as usual. CrewBase opens one adapter for the whole crew and closes it after kickoff().

What comes back

The first call above printed Example Domain 958 chars of markdown. Here's what each tool hands back through the adapter:

Inside a crew the agent sees the same strings, and the backstory tells it to follow up with resource_fetch. Two more things: meta.status_code isn't reliable (when we fetched a page that answers 403, status said success and status_code was missing), and a fresh fetch takes 10 to 40 seconds (19 in this test), against under half a second for a page fetched recently.

On 29 Oct 2026 we re-ran the direct calls in both examples with the key in the header, using crewai and crewai-tools 1.15.33 and mcp 1.30.0 on Python 3.12, and called crawl_fetch through the mcps field with the header; the 30-second limit is MCP_TOOL_EXECUTION_TIMEOUT in that version's source. Before that, on 25 Oct with 1.15.30 and the key in the URL, we ran both examples against the live UpRock server. The direct calls returned example.com's title and markdown. For the crew we swapped LLM(model=MODEL) for a scripted test model with no API key: it called crawl_fetch and then resource_fetch through UpRock and finished with the page title. The @CrewBase class ran the same way with a minimal agents.yaml. The key-in-URL mcps names come from that version, with a made-up key and a model that recorded what it was sent.