Hermes Agent × Browser Use
The web now runs through one tool.
Hermes replaced 12 browser tools with one call. Its model writes Python; Browser Use runs it against a local or Cloud browser.
Hermes's browser guide: “This is the default browser mode” — when no backend is set and the CLI can run, the agent gets the single browser_exec tool.
Plain English in. Python at the tool boundary.
Hermes's tool-calling model writes Python into browser_exec. Browser Use executes that code and reaches the browser over CDP.
Before · 12 tools in every prompt
After · 1 tool
browser_exec
The model writes Python; Browser Use executes it against whichever browser is configured.
browser_exec(code="""
new_tab("https://example.com")
click_at_xy(240, 380)
""")Same tasks solved. Two thirds fewer tokens.
Twelve tool definitions sat in every prompt. One does now, and the model stopped chaining calls to do one thing.
| Model | Browser tools | Solved | Mean tokens | Tool calls | |
|---|---|---|---|---|---|
| opus-4.8 | 12 built-in tools | 18/18 | 64,594 | 4.1 | |
| opus-4.8 | browser_exec | 18/18 | 25,934 | 2.0 | −60% |
| kimi-k3 | 12 built-in tools | 18/18 | 56,464 | 5.3 | |
| kimi-k3 | browser_exec | 18/18 | 19,230 | 2.4 | −66% |
6 multi-step web tasks × 3 reps per arm, web-fetch tools stripped so every arm had to drive the browser. Both arms solved every task; the tokens fell at accuracy parity.
Measured in Hermes's merged pull request across 204 live runs. Read the benchmark.
One interface. Two places to run the browser.
Use a local browser when the agent needs your open tabs, cookies, or a browser you can watch.
Use a Cloud browser when the task needs managed sessions, profiles, proxies, or parallel browsers.
Any general agent that can call the tool can use the same browser layer.