Getting it into your agent
It runs from inside its repository, so the clone comes first — what it calls does not travel with the file alone.
git clone --depth 1 https://github.com/lexmount/browseruse-agent-benchnpx agentmods add skills/lexmount/browseruse-agent-bench/bubench-runWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/lexmount/browseruse-agent-bench/bubench-run)<a href="https://agentmods.dev/skills/lexmount/browseruse-agent-bench/bubench-run"><img src="https://agentmods.dev/badge/skills/lexmount/browseruse-agent-bench/bubench-run/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/lexmount/browseruse-agent-bench/bubench-run"><img src="https://agentmods.dev/badge/skills/lexmount/browseruse-agent-bench/bubench-run.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00166 | $0.01729 |
| Opus 5.5 | $0.00066 | $0.00692 |
| Sonnet 5.5 | $0.00033 | $0.00346 |
| Haiku 4.5 | $0.00017 | $0.00173 |
Grade A, and why
bubench-run scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 1mo ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 151 lines — stays where its author put it; the contents beside it link to each section on GitHub.
BrowserUse-Bench Run & Eval Workflow
This skill guides running benchmark experiments on browseruse-bench: launching agent runs, evaluating results, and chaining multiple models in sequence — all using background processes and monitors so Claude stays responsive while jobs run.
Key Commands
# Run a model on specific tasks (background)
uv run scripts/run.py --agent browser-use --data LexBench-Browser \
--mode specific --model-name <model-name> \
--task-ids <ids...> \
> output/logs/run/<label>.log 2>&1 &
# Evaluate results (background)
uv run scripts/eval.py --agent browser-use --data LexBench-Browser \
--model-id <model-id> \
> output/logs/eval/<label>.log 2>&1 &
# Evaluate a specific timestamp directory (for incremental/second passes)
uv run scripts/eval.py --agent browser-use --data LexBench-Browser \
--model-id <model-id> --timestamp <YYYYMMDD_HHMMSS> \
> output/logs/eval/<label>.log 2>&1 &
Model Name → Model ID Mapping (from config.yaml)
--model-name |
--model-id (in experiments path) |
|---|---|
deepseek |
deepseek-v4-pro |
minimax |
MiniMax-M2.7 |
claude |
dmx-claude-opus-4-7 |
gemini |
gemini-2.5-pro |
gpt |
gpt-5.5 |
Always check config.yaml to confirm the current model_id before running eval — it must match the directory name under experiments/.
Output Paths
experiments/LexBench-Browser/All/browser-use/<model-id>/<timestamp>/
tasks/ # one subdir per task
tasks_eval_result/
eval.log # live eval progress
task_gpt-5.4_per_task_threshold_stepwise_summary.json
Run logs: output/logs/run/<timestamp>.log
Eval logs: output/logs/eval/<label>.log
Monitoring Pattern
Always use Monitor (not polling loops) to track background jobs.
Run progress (persistent — tasks take hours):
tail -f output/logs/run/<label>.log | grep --line-buffered \
-E "\[[0-9]+/N\]\[[0-9]+\] (completed|failed)|Run complete|ERROR"
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 1mo ago First seen · 151 lines · 166 tokens per session scan A ba9f0bfa20a7
bubench-run is a skill published in the GitHub repository lexmount/browseruse-agent-bench (20 stars, last pushed 2mo ago), licensed Apache-2.0. It adds 166 tokens to every session and 1,729 once invoked, about $0.0007 per session on Opus 5.5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
Accessibility Tree Inspection
Inspect accessible names, roles, and tree structure to drive or audit browser UI behavior.
Browser Smoke Testing
Open a local web target and verify that the main page renders and core controls respond.
Form Interaction Flow
Drive browser forms and controls through realistic user interactions.
macOS Computer Use
Operate macOS applications, windows, files, and desktop workflows.
browser-visual-review
Use when reviewing local browser screenshots for layout shifts, visual regressions, and viewport state.
web-ui-test
Test the IronClaw web UI using the Claude for Chrome browser extension.