Hermes Camofox + Which AI Models To Run It With: Real GoldieBench Scores For Every Browser Job

Hermes Camofox gives your Hermes agent a stealth Firefox to browse with, but the model you put behind it decides how well the agent reads each page and picks what to click.

Camofox is the official Hermes plugin that drives Camoufox, a Firefox fork with C++-level fingerprint spoofing, through a local server.

Below I match models to Camofox jobs using only scores I read on live GoldieBench pages on 31 October 2026, each quoted with its page.

If you want my team to help you plan the agent side of your marketing, book a free SEO strategy session.

Camofox uses two model slots, not one

The main Hermes model drives the ten Camofox tools: it reads browser_snapshot output and decides on browser_click, browser_type and the rest.

A second slot powers browser_vision, which per the plugin README always answers through the auxiliary vision model.

That model is configured under auxiliary.camofox_vision.* and inherits auxiliary.vision unless you override it.

So the pairing question has two answers: a reasoning model for the loop, and a vision model for screenshots.

What GoldieBench scores measure, and what they don't

The GoldieBench methodology page is explicit: every model gets the same fixed prompt, one shot, and must return one self-contained HTML file.

Julian scores each result from 0 to 10, and the leaderboard ranks models by their average across the tasks they were scored on.

That's a test of building, not browsing.

Building still leans on the same skills a browsing agent uses: following instructions exactly, holding a plan across many steps, and not losing track halfway.

No GoldieBench task drives a browser through Camofox, so read the scores as a signal of general coding and reasoning strength, not of click accuracy.

Where a model page adds something closer to agent work, like context recall or long runs, I call it out.

The scores, as read on goldiebench.com

The leaderboard counted 33 models, 50 tasks and 1310 live demos when I read it.

Kimi K3 for long, unattended Camofox runs

Kimi K3 scores 7.89 across 50 tasks on its model page.

The number that matters more for browsing is on the same page: a needle test with exact recall from 162k tokens of noise in 18s.

A Camofox session piles up snapshot after snapshot, and the plugin saves any snapshot over the 15000 threshold to a file the agent pages through, so recall over a long context pays off.

The page also says Julian wires it in as the kimi-k3 Hermes profile for long unattended agent runs.

Its Terminal Bench note comes from launch-day community benchmarks, which the page labels as external, not GoldieBench medal scores.

Claude Opus 5 for the hardest browser flows

Claude Opus 5 sits at 8.27 over 50 tasks, the top single-vendor model on the leaderboard I read.

Its page lists "hardest agentic builds" and whole-repo reasoning under best for, with a 1,000,000-token context.

Use it for multi-step flows where one wrong click wastes the run, like a checkout you approve at the end or a portal with nested menus.

Its own trade-off line warns that reasoning-by-default eats token budgets, so don't make it the default for simple reads.

Gemini 3.6 Flash for high-volume page reads

Gemini 3.6 Flash scores 7.08 on 50 tasks, well below the flagships.

Its page still names it for high-volume agentic work where token cost dominates, and for routing the everyday 90% while a flagship takes the hard 10%.

That fits Camofox jobs that are mostly reading: open a dashboard, snapshot it, summarize it.

The 1,000,000-token context on the same page leaves plenty of room for long snapshots.

Why Hermes MoA is the wrong lane for click-by-click browsing

Hermes MoA scores 8.17 over 47 tasks, third on the leaderboard.

Its model page explains the catch: a panel of models drafts in parallel and a chair merges them, at about 110–140s per single-file build.

A Camofox loop makes many small decisions in a row, and that panel latency would land on every one.

Keep MoA for the report you write after the browsing is done, not for the browsing itself.

Grok when the research is about X

Grok averages 8.09 on the leaderboard, with 43 of 47 tasks scored.

Its page lists real-time access to X timeline data as the unique strength, and a 256,000-token context.

The same page flags that API access is gated behind X Premium, which it calls awkward for backend agent loops.

So pick Grok for X-heavy research, and a different model for the Camofox driving.

Matching a model to each Camofox job

Here's how I'd map the scores above onto the work a Camofox agent actually does.

Start every new job on the cheapest lane that can do it, and move up only when the run keeps failing.

Log which model ran which job for a week, then keep the pairings that finished cleanly.

Setting the models in Hermes

Pick the main model with hermes model.

Point browser_vision at a model of its own under auxiliary.camofox_vision.*, or leave it inheriting auxiliary.vision.

Then install the browser side: hermes plugins install camofox, with a Camofox server from jo-inc/camofox-browser running on port 9377.

The Hermes browser docs note that older Camofox setups from core migrate to the plugin automatically on update.

For logins, the vault fill on Camofox tabs landed on 28 October 2026 in commit 0a93a234 and commit 6f7d49b5, and no tagged release has it yet; the latest tag is v0.22.5 from 23 October 2026.

The credential vault docs cover how fills stay bound to the exact site origin.

The pairing I'd start with

• Daily reads: Gemini 3.6 Flash as the main model.
• Long or tricky runs: switch the profile to Kimi K3 or Claude Opus 5.
• The write-up afterward: hand the collected notes to Hermes MoA.

Re-check the scores on the models page before you commit, because the leaderboard moves as new models land.

Want the profiles and prompts for this split set up for you? Join the AI Profit Boardroom.

Also On Our Network

• Agent Operatorsthe six-step Camofox setup, from npm start to persistent logins
• AI Income Deskwhether a small business should run Camofox, with three real use cases
• AI Tool Verdicta nine-point scorecard rating Camofox from setup to vault fills
• agentos.guidefive copy-paste prompts that get your agent running Camofox
• aiprofitboardroom.comthe complete Hermes Camofox guide on the AI Profit Boardroom blog

Start daily reads on Gemini 3.6 Flash, move hard runs to Kimi K3 or Claude Opus 5, and if you want help, book a free SEO strategy session or join the AI Profit Boardroom to run them with Hermes Camofox.