⭐ Get the Agent OS + join 3,400+ founders inside the AI Profit Boardroom → Join AIPB

GoldieBench Blog · 6 min read

Grok Bot Custom MCP + Which AI Models Should Write Your Server And Build On Its Data, With Real GoldieBench Scores

Grok Bot picks its own model, so the scores matter for the server code you write and the pages you build from its data.

Illustrated benchmark leaderboard of glowing gold bars wired to a plug-in server block on a dark navy background

Grok Bot custom MCP lets you plug your own server into Grok Bot, but it doesn't let you pick the model that calls it, so the model question moves to the code and the output around the server.

The Grok Bot FAQ says Grok Bot picks the model for each task, there is no model picker, and the choice can differ from one request to the next.

So this page answers two narrower questions: which model should help you write a command server, and which model should build what you make from its data.

Every score below I read on live GoldieBench pages on 12 November 2026, and each one names the page.

Want this mapped to your own tools? Book a free SEO strategy session with my team.

Custom MCP in one table

TypeWhat you enterWhere Bots reach itNetwork allowlist check
Remote HTTPSServer URL, optional HTTP headers, optional OAuth Client IDCursor's serversOnly with Apply to MCP servers on
CommandCommand and ArgumentsThe Bot's computerAlready under the network policy

Both types are added under Add a custom connector in Connect apps, per the computer and apps docs, or by asking a Bot to add the server by URL or command.

The network rules come from the security page.

Neither row involves choosing a model.

What GoldieBench measures, and what it can't tell you here

The GoldieBench home page gives every model the same fixed prompt set: one shot, one HTML file out, scored 0 to 10.

On 12 November 2026 it listed 1310 live demos, 50 tasks, 33 models and 1041 curated verdicts.

That is a build test: games, sims, visuals and pages.

It doesn't score MCP tool calls, sign-in handling or server code.

So treat every number here as evidence about one-shot build quality, which matters for what you do with server data, not for the connection itself.

Grok: 8.09 average, rank #7

MetricGrokPage I read
Average score8.09/10Grok model page
Tasks47 tested, 43 scoredGrok model page and leaderboard
Medals5 gold, 1 silver, 1 bronzeGrok model page
Leaderboard rank#7Leaderboard home
Context256,000 tokensGrok model page

The Grok model page lists its best use as workflows that need live X context.

Its top build was Twilightvale at 9.5, with Game, Rpg, Landing Page and Webos at 9.0 each.

Landing Page and Webos are page builds, the category closest to a dashboard or report you'd make from server data.

Its weakest page build on the same scorecard was Aipbpromo at 7.0, so page quality isn't uniform.

Grok 4.7: 7.15 average, and the Terminal-Bench number to watch

The Grok 4.7 model page shows 7.15/10 across 20 scored tasks, with 2 silver and 3 bronze medals and a 500,000-token context.

The leaderboard home ranks it #17.

Its best one-shots were Racing and Neoncity at 8.7, then Dragonrealm and Arcade at 8.6.

The same page lists xAI-reported coding numbers: Terminal-Bench 4.0 at 38.0%, DeepSWE v1.1 (high effort) at 71.0% and CursorBench 4.0 at 46.3%.

The page calls Terminal-Bench the weak spot xAI reports itself.

That's the figure I'd look at for command servers, because a Command connector is a program started from the command line.

The top of the board for code and builds you route elsewhere

RankModelAvg /10ScoredBest at
#1Fusion8.5947/47Games
#2Claude Opus 58.2750/50Sims
#3Hermes MoA8.1747/47Visuals
#4GPT-5.6 Sol8.1650/50Pages
#5Claude Fable 58.1047/47Games
#7Grok8.0943/47Games
#11GLM-5.27.7747/47Pages

Every row is from the goldiebench.com leaderboard on 12 November 2026.

The home page says Fusion is currently graded by Claude's AI judge, so treat its lead as provisional.

For page-style output, the two models the board marks as best at Pages are GPT-5.6 Sol and GLM-5.2.

Writing a command server: what the scores do and don't say

A Command connector runs a program you supply, from /workspace or wherever you keep it.

Writing that program is ordinary coding, and you can use any model you like for it outside Grok Bot.

GoldieBench doesn't score server code, so no number here tells you which model writes the best MCP server.

What it does show is which models hold a long single-file spec together in one shot, which is a reasonable proxy for writing a small, self-contained server script.

Whichever model writes it, test the server read-only through Grok Bot before any write task.

Keep the script small enough to read in one sitting, because you are the one approving what it can touch.

Ask the model to list every tool the server exposes and what each one changes, then check that list against what you actually need.

Anything that writes, sends or deletes should be a separate tool you can leave out of the first version.

Two docs facts change how you'd brief that model.

First, the docs say virtual environments and node_modules aren't saved, so ask for a server with few dependencies and a reinstall note.

Second, a command server is available to your other Bots too, so ask for read-only tools first and write tools only where you need them.

Building on server data: where page scores matter

Most custom servers end up feeding a report, a dashboard or a page.

Leave the questions inside Grok Bot, where the server already lives: today's bookings, open jobs, low stock.

Export the data when you want a built thing, and route that build to a model you choose from the board.

Grok's 9.0 on Landing Page and Webos makes it a credible choice for that step too.

The same split works for catalog apps; our Grok Bot connectors model guide covers that side.

Rules that apply whatever model runs

  • Only add MCP servers you trust; a connected app can do whatever the authorised account can.
  • OAuth tokens live with Cursor's connector backend; Bots use the tools, never the tokens.
  • A team's MCP policy can block a server; it then shows Disabled by team admin.
  • On Enterprise, turn on Apply to MCP servers to hold remote servers to the allowlist; it's off by default.
  • The changelog release of 11 November 2026 marks custom connectors on Team Bots as shared with everyone.

My pairing plan

  1. Add the server in Grok Bot and authenticate it; don't worry about the model there.
  2. Write or maintain any command server with a model you pick, briefed for few dependencies and read-only tools first.
  3. Ask Grok Bot your day-to-day questions through the server with @.
  4. Route page and dashboard builds to a model from the board, checking its Pages scores first.
  5. Re-read the scores before each big build; the board changes as models ship.

Bring your shortlist to a coaching call in the AI Profit Boardroom and we'll compare notes.

Also On Our Network

Keep the connection in Grok Bot, put the code and the builds on models the scores favour, and book a free strategy session if you want help pairing models with your grok bot custom mcp setup.

FAQ

Can I choose which model calls my custom MCP server?

No. The Grok Bot FAQ says it picks the model for each task, there is no model picker, and the choice can differ between requests.

Does GoldieBench test MCP servers?

No. It scores one-shot HTML builds from a fixed prompt set. Use it to pick a model for building with server data, not for the connection.

What does Grok 4.7 score on Terminal-Bench?

The Grok 4.7 model page lists 38.0% on Terminal-Bench 4.0, a number xAI reports itself and the page calls the weak spot.

Which GoldieBench models are best at Pages?

On 12 November 2026 the leaderboard marked GPT-5.6 Sol (8.16, #4) and GLM-5.2 (7.77, #11) as best at Pages.

The same stack Julian uses

Run this stack yourself.

Every demo on this bench was built inside the Agent Operating System — one prompt, one shot, single HTML file out. The Agent OS, the prompts, the templates, the weekly walkthroughs and 3,400+ founders shipping with it every day all live inside the AI Profit Boardroom.

3,400+founders
258documented wins
38countries
4coaching calls a week