HyperTeamsSame model, same 400 tasks.
Only the browser changed.
Seven browser tools — the hands an agent uses on the web — each driven by the same Claude Code and the same model through every task. The HyperTeams built-in browser is first in overall score and overall accuracy, and the fastest to finish.
- Overall score 84.9
- Overall accuracy 83.9%
- Real services (REAL) 74%
- Web interaction (MiniWoB++) 95%
- #1HyperTeamsHyperTeams84.983.9%
- #2AsideAside83.982.2%
- #3Ego LiteEgo Lite81.980.4%
- #4Playwright MCPPlaywright MCP81.680.9%
- #5agent-browseragent-browser80.880.6%
- #6DevTools MCPChrome DevTools MCP80.380.1%
- #7browser-usebrowser-use80.280.1%
- model — Claude Sonnet 5 for all seven tools
- 1
- tasks — four public benchmarks plus 8 prompt-injection scenes
- 400
- browser tools — same prompt, same time limit
- 7
- LLM judges — every task graded by a program
- 0
Results by benchmark
Cell shade is the pass rate — darker means more solved.
Swipe sideways to see all seven tools
| Benchmark | HyperTeams | Aside | Ego Lite | Playwright MCP | DevTools MCP | browser-use | agent-browser |
|---|---|---|---|---|---|---|---|
| Overallmean of the four pass rates | 84.9 | 83.9 | 81.9 | 81.6 | 80.3 | 80.2 | 80.8 |
| Overall accuracyshare of 392 tasks solved | 83.9% | 82.2% | 80.4% | 80.9% | 80.1% | 80.1% | 80.6% |
| REALRead the original benchmark (opens in a new tab)121 tasks | 74%89/12195% CI 65–81 | 69%83/12195% CI 60–76 | 69%83/12195% CI 60–76 | 70%85/12195% CI 62–78 | 69%83/12195% CI 60–76 | 70%85/12195% CI 62–78 | 70%85/12195% CI 62–78 |
| WebGamesRead the original benchmark (opens in a new tab)53 tasks | 96%51/5395% CI 87–99 | 98%52/5395% CI 90–100 | 93%49/5395% CI 82–97 | 91%48/5395% CI 80–96 | 87%46/5395% CI 75–94 | 87%46/5395% CI 75–94 | 87%46/5395% CI 75–94 |
| MiniWoB++Read the original benchmark (opens in a new tab)128 tasks | 95%122/12895% CI 90–98 | 92%118/12895% CI 86–96 | 88%112/12895% CI 81–92 | 92%118/12895% CI 86–96 | 94%120/12895% CI 88–97 | 94%120/12895% CI 88–97 | 94%120/12895% CI 88–97 |
| GAIA webRead the original benchmark (opens in a new tab)90 tasks | 74%67/9095% CI 65–82 | 77%69/9095% CI 67–84 | 79%71/9095% CI 69–86 | 73%66/9095% CI 63–81 | 72%65/9095% CI 62–80 | 70%63/9095% CI 60–79 | 72%65/9095% CI 62–80 |
74%Also finished the most real-service jobs
Mail, shopping, restaurant booking, calendars and hiring on 11 clones of real services, graded by the state each site is left in. HyperTeams placed first with 89/121.
Paired task by task, it solved more than every one of the six other tools.
Same tasks — solved by only one side
HyperTeams onlyRival only- Aside82
- Ego Lite71
- Playwright MCP51
- Chrome DevTools MCP93
- browser-use62
- agent-browser51
Tasks both solved or both missed are not counted (exact McNemar test).
95%Solved the most when the page has to be operated directly
Forms, lists, date pickers and drag-and-drop, graded by the page itself. HyperTeams placed first with 122/128.
Paired task by task, it solved significantly more than Ego Lite.
Same tasks — solved by only one side
HyperTeams onlyRival only- Aside73
- Ego Litesignificant · p 0.031144
- Playwright MCP40
- Chrome DevTools MCP64
- browser-use53
- agent-browser31
Tasks both solved or both missed are not counted (exact McNemar test).
MiniWoB++
11s ahead of the next fastest- HyperTeams 25s
- Playwright MCP 36s
- DevTools MCP 69s
- Aside 79s
- agent-browser 93s
- Ego Lite 107s
- browser-use 114s
REAL
- HyperTeams 85s
- DevTools MCP 85s
- agent-browser 90s
- Playwright MCP 92s
- browser-use 108s
- Aside 135s
- Ego Lite 190s
Finishes fastest
Median time of passing runs. HyperTeams was the fastest on MiniWoB++ and among the fastest on REAL.
Each dot is one tool.
When the agent gets stuck, where can you take over?
Hand off hours of work and people walk away. Scores may be close, but where you can step in differs by tool.
- Phone
- Another computer
- The computer the agent runs on
Anywhere
Watch live from your phone or another computer's dashboard, and take over on the spot.
- Phone — yes
- Another computer — yes
- The computer the agent runs on — yes
- HyperTeams
Only at that computer
You have to sit at the computer the agent runs on to watch or take over.
- Phone — no
- Another computer — no
- The computer the agent runs on — yes
- Aside
- Ego Lite
Can't watch
No way for a person to see the agent's screen or take over.
- Phone — no
- Another computer — no
- The computer the agent runs on — no
- Playwright MCP
- Chrome DevTools MCP
- browser-use
- agent-browser
How we measured
- Same conditions
- Claude Code · claude-sonnet-5 · same prompt · same time limit. Only the browser layer differs. Apple M1 · 16GB, Sep 28, 2026 – Oct 7, 2026.
- Program grading
- Exact match, the page's own score, or the state left on the site. No LLM judges; differences between tools use exact McNemar tests on paired tasks (p < 0.05).
- Complete
- Seven tools × 400 tasks, none skipped. Timed-out and failed runs were measured once more, by the same rule for all seven tools.
Versions measured
| Tool | Version |
|---|---|
| HyperTeams | 2026-10 |
| Aside | 1.0.922.1 · 1.0.928.1 |
| Ego Lite | 0.5.1.13 |
| Playwright MCP | @playwright/mcp 0.0.82 |
| Chrome DevTools MCP | chrome-devtools-mcp 1.10.1 |
| browser-use | browser-harness 0.1.13 |
| agent-browser | 0.38.1 |
Hand it a job yourself
The built-in browser ships with HyperTeams. Turn on browser tools in a folder's settings and every task there — Claude Code, Codex, Antigravity or Pi — uses the web through the same tools.
curl -fsSL https://hyperteams.net/install.sh | bash