Gemini 4 Argon MCP: live web and company data

Gemini 4 Argon MCP gives Google's new model data it cannot see on its own: a live web search, a page read as markdown, a company profile, Google results with positions. Argon went generally available in the Gemini API on 2026-10-13, and Google's API docs describe remote MCP servers for the Interactions API.

This post shows the setup Google documents, three data tasks to run once it works, and when Google's own search grounding is the cheaper choice. You need a paid Gemini API key, because Argon is not on the free tier. By the end, Argon calls one MCP server that reaches many data providers, and you know what each search costs before it runs.

Last checked 2026-10-13, against Google's pages and looot's live catalog.

Gemini 4 Argon tools: what does it support?

Function calling, search grounding, URL context, code execution and file search are all listed as supported on Google's Gemini 4 Argon model page (read 2026-10-13). The model id is `gemini-4-argon`, with a 1,048,576 token context window shared by input and output.

Google's launch post (read 2026-10-13) shows Argon driving a Blender MCP server "in an autonomous loop". It mentions MCP only in its Blender demos. The setup below comes from the API docs, not from the launch post.

How do you connect an MCP server to Gemini 4 Argon?

You pass the server as a tool in an Interactions API call. Google's function calling docs (read 2026-10-13) define a tool of type `mcp_server` with a `name`, a `url`, optional `headers` sent on every request, and optional `allowed_tools`. The server must speak Streamable HTTP, and its name must not contain a hyphen.

looot's server is `https://api.looot.ai/mcp`, Streamable HTTP. For a script with no browser, create an agent token in looot settings and keep it in an environment variable. Then:

pip install google-genai
export GEMINI_API_KEY=...   # paid tier key
export LOOOT_TOKEN=...      # agent token from looot settings

import os
from google import genai

client = genai.Client()
interaction = client.interactions.create(
    model="gemini-4-argon",
    input="Search the web for the Gemini 4 Argon launch post and read it as markdown.",
    tools=[
        {
            "type": "mcp_server",
            "name": "looot",
            "url": "https://api.looot.ai/mcp",
            "headers": {"Authorization": f"Bearer {os.environ['LOOOT_TOKEN']}"},
            "allowed_tools": ["search", "inspect", "run", "balance"],
        }
    ],
)
print(interaction.output_text)

The shape follows Google's Python example, which uses `gemini-3.8-flash`; we changed the model id, the prompt and the server, and added `headers`, `allowed_tools` and a `print`. We have not run this exact call against Argon, so run it once with a cheap prompt first. `allowed_tools` keeps the model on the four tools it needs: find a job, read its inputs and price, run it, check the balance.

Three data tasks to give Argon once it is connected

Each task is one looot job. The model searches the catalog, reads the price, then runs the job with fallback on, so a miss at one provider moves to the next.

1. Search the web. The job web.search returns titles, links and snippets from providers such as Brave, Exa and Tavily.
2. Read a page as markdown. The job web.scrape.markdown turns any URL into clean text the model can quote.
3. Profile a company. The job company.enrich takes a domain and returns the name, size and links. For SEO work, google.serp.organic returns Google results with their positions.

We ran two of these through looot on 2026-10-13 to see what comes back. A Google search for "gemini 4 argon mcp" returned seven results, with Google's launch post at position 6 and n8n's Argon workflow guide at position 3. A company lookup for earendil.com missed at the first provider, which billed nothing, and the second provider found the company with a size of 11 to 50 people. The run record shows both attempts and which one was charged.

Gemini 4 Argon MCP or search grounding?

Use grounding when you want an answer backed by Google Search and you stay inside the free 5,000 requests a month. Use an MCP data server when you need records grounding does not return: a company profile, a work email, ranked Google positions, or a choice of search provider. Many agents will use both. If you only need one search vendor, a single-vendor server such as Exa's MCP server can sit in the same `tools` list.

Grounding costs $14 per 1,000 requests after the free allowance (Gemini API pricing, read 2026-10-13). An external search tool is billed per call by its provider, and the table above shows what each one charges today. One prompt can trigger several Google searches, and each one is billed. Compare the two at your real volume, not at one call.

For reading one known page, Google's own URL context tool is the cheaper start: it is free as a tool and only the text counts as input tokens. A scraper earns its place when a page blocks plain fetches or needs JavaScript to render.

What it cannot do

• It does not change what Argon costs. Your Gemini tokens, including the text the tools return, are billed by Google at the rates above.
• It cannot sign in to sites for you, read private accounts or post on social media.
• It does not run on a timer. A weekly search or an alert needs scheduled runs, not available yet.
• A miss is a normal answer. If no provider has the company or the email, the run says so instead of guessing.
• The same server works in other agents. The Pi coding agent setup uses it with one command.
• We have not tested Gemini CLI with Argon. looot's Gemini CLI setup works for the models the CLI offers, and the Interactions API above is the path Google documents for remote MCP.