Browser agents make a model drive the screen, click by click. Protocolis compiles any site instead, with no model inference and no site-specific setup, into verified operations the agent calls by name. First on completion, cost and speed across more than 2,800 runs: a small model on Protocolis beat GPT-6.1 Sol on Browser Use, 28.6 seconds and correct against 97.2 seconds and wrong.
Protocollum ex ipsa applicatione
Every other agent makes you shift gears yourself. Browser Use ships its own agent, a flash mode, a tools mode and a hosted Ultrafast; the faster the mode, the fewer tasks it finishes. So you pick a full mode for hard tasks and a fast one for easy ones, and pay for the wrong choice either way. Protocolis Adaptive is one mode that shifts on its own: step by step, the engine’s evidence decides whether a small choice model is enough or the agent’s model is needed. No fast mode, no full mode, and no trade. Measured against Browser Use, Browserbase’s Stagehand, Microsoft’s Playwright MCP and Jev Ultrafast on the same model and tasks.
Every screenshot stack holds in one corner and falls away in the others: a tools mode for small models, an agent for frontier ones, a fast mode for short tasks. That is why Browser Use ships six ways to run the same agent. One compiled layer holds across the range, and the model above it can be whatever the budget allows.
Each point is one system at one model size (Small, Strong, Frontier), taken from the detailed benchmarks below. Up is more tasks completed; left is less money per completed task, and failed runs count against it. Depth is time per completed task: the further back a point's faint line runs, the slower it is. Pick a rival in the tables and its point is drawn bright here.
Each point is a row from the detailed benchmarks below; a system with no figure in a cell has no point there. Pick a rival in the tables and its line is drawn bright here. Long tasks are being re-measured.
Replayed from the saved runs, shown at 2× speed; the clocks show real seconds. The panes call Adaptive by its internal name, Auto. Every answer checked against the final page and a key written before the run.
Replayed from the saved runs, shown at 2× speed; the clocks show real seconds. Every answer checked against the final page and a key written before the run.
Replayed from the saved runs, shown at 2× speed; the clocks show real seconds. Every answer checked against the final page and a key written before the run.
Adaptive’s choice model made 2 picker calls here, counted in its cost. Replayed from the saved runs of one four-arm race, shown at 2× speed; the clocks show real seconds. The Adaptive pane carries its internal name, Auto. Every answer was checked against the final page and a key written before the run. Selected runs.
Adaptive’s choice model made 3 picker calls here, counted in its cost. Replayed from the saved runs of one four-arm race, shown at 2× speed; the clocks show real seconds. The Adaptive pane carries its internal name, Auto. Every answer was checked against the final page and a key written before the run. Selected runs.
Adaptive’s choice model made 3 picker calls here, counted in its cost. Replayed from the saved runs of one four-arm race, shown at 2× speed; the clocks show real seconds. The Adaptive pane carries its internal name, Auto. Every answer was checked against the final page and a key written before the run. Selected runs.
About 1/118 of the cost. Shown at 8× speed.
Browser Use answered from Wiktionary instead of Cambridge.
About 1/22 of the cost.
About 1/235 of the cost.
About 1/149 of the cost.
About 1/171 of the cost.
About 1/129 of the cost.
About 1/210 of the cost.
Browser Use faster here; Protocolis at about 1/117 of the cost.
Same answer.
Browser Use’s own eval setup.
Same answer, 7.5095 × 10^33.
Browser Use reached the same answer but ended without a final page, so the judge failed it. Shown at 8× speed.
Shown at 8× speed.
Shown at 8× speed.
Browser Use faster here.
Shown at 3.5× speed.
Take 1 of the 2 usable Apache takes, out of 4 recorded. The other two had a blank screen capture and a wrong call count on the board. One task, not a benchmark result. BU Ultrafast ran as Browser Use sells it, on their cloud browser and model through their API; Protocolis Adaptive ran on one Mac. Both started on the same second, each clock runs from GO to that agent’s finish, and the board names each model. Across all 8 takes of these two tasks, Adaptive finished first every time and BU Ultrafast answered correctly every time. Shown in real time.
Take 2 of the 2 usable Gutenberg takes, out of 4 recorded. The other two showed BU Ultrafast as failed on the board when it had answered correctly. One task, not a benchmark result. BU Ultrafast ran as Browser Use sells it, on their cloud browser and model through their API; Protocolis Adaptive ran on one Mac. Both started on the same second, each clock runs from GO to that agent’s finish, and the board names each model. Across all 8 takes of these two tasks, Adaptive finished first every time and BU Ultrafast answered correctly every time. Shown in real time.
BU Ultrafast called through Browser Use’s public API, as their page shows. Over six runs it once answered “Virginia Tech, 1872” and once skipped the links for a web search; Protocolis followed the links to 1802 every time. Shown in real time.
Shown in real time.
Two page hops, then a read. Jev stayed on Collins’s page and stopped at its own 120-call budget; a second take went the same way (Protocolis 6.2 s). Shown in real time.
Browser Use tools mode faster here.
Same answer.
Same answer.
Same coding agent, model and prompt; fresh browsers on one Mac. Answers checked against site data or Browser Use’s judge.
Matched model, task and prompt; fresh browsers on one Mac. Answers checked against a prewritten key or Browser Use’s judge.
Matched model, task and prompt; fresh browsers on one Mac. Answers checked against prewritten keys.
Grok 4.6 in every arm, on the same task and cloud worker. Answers checked against the final page and prewritten keys.
A smaller Protocolis model against a frontier Browser Use model, with the same task and budget. Answers checked against prewritten keys or Browser Use’s judge.
Answers checked against prewritten keys.
Passwords stay in the broker and never reach the agent’s model.
Standard vs Browser Use’s own agent · GLM-5.3-flash · 98 held-out tasks
Ultrafast: Adaptive (lean) vs Jev Ultrafast · same 40 runs, matched clock · 40/40 vs 26/40 completed · Jev 4.2 s vs 8.8 s per completed task
Each cell is the share of tasks completed, brighter for more; a dash means not measured. Lowest is a system's weakest measured cell; Cells is how many of the nine it was measured in. † the test kit ran with challenge handling off. Point at a cell for its numbers; pick it to open that cell in the detailed tables.
Sixteen capabilities, compared product by product.
| Capability | Protocolis | Browser UseOwn agent | Browser UseOpenAI tools | Browser HarnessBrowser Use | Stagehandv4 docs | Playwright MCPMicrosoft | UnbrowseHosted service | Browser Controlkeon |
|---|---|---|---|---|---|---|---|---|
| ★Just-in-time cold compiler engine | ✓ | — | — | — | — | — | — | — |
| ★Automatic transmission: one mode that shifts step by step | ✓ | — | — | — | Partial | — | Partial | — |
| ★Self-healing: verification that learns from outcomes | ✓ | Partial | — | — | Partial | — | Partial | — |
| ★Application operations compiled on first contact | ✓ | — | — | — | — | — | — | — |
| ★Compilation without model calls | ✓ | — | — | — | — | — | — | — |
| ★Action effects verified on first contact | ✓ | Partial | Partial | Partial | Partial | Partial | — | Partial |
| ★Task completion verified on first contact | ✓ | Partial | Partial | Partial | Partial | Partial | — | Partial |
| ★Compiled knowledge reused across visits | ✓ | — | — | — | Partial | — | ✓ | — |
| ★Verified replay with protected-action stops | ✓ | Partial | — | — | Partial | — | Partial | — |
| Screenshot-free execution and result checks | ✓ | Partial | Partial | ✓ | ✓ | ✓ | — | ✓ |
| Built-in bot-protection handling | ✓ | Partial | Partial | Partial | Partial | — | ✓ | Partial |
| Challenge handling and isolated escalation | ✓ | Partial | Partial | Partial | Partial | — | Partial | Partial |
| Challenge lessons shared across sites | ✓ | — | — | — | — | — | — | — |
| Model-blind credential broker | ||||||||
| ★Credentials bound to verified fields | ✓ | Partial | — | — | Partial | Partial | Partial | — |
| ★New API keys captured model-blind | ✓ | — | — | — | — | — | — | — |
| Passkeys | ✓ | — | — | — | — | — | — | — |
| Full matchesOut of 16 listed capabilities | 16 / 16 | 0 / 168 partial | 0 / 165 partial | 1 / 164 partial | 1 / 169 partial | 1 / 163 partial | 2 / 165 partial | 1 / 164 partial |
✓ Matches the defined capabilityPartial: related support with limits— Not established in reviewed sourcesPatent pending
Documentation review · 5 November 2026 · Ratings use the definitions and scope below; they are not new benchmark results.
Each alternative was reviewed individually against the same definitions. “Partial” includes narrower mechanisms and separately supplied cloud services. A dash means the reviewed sources do not establish the defined capability, not that support is impossible. Related technology does not earn partial credit when it fails an explicit cold or first-contact requirement. Protocolis ratings describe our engine’s supported capabilities; competitor ratings describe their published documentation and source, not independent runtime testing. Current documentation may differ from the versions used in the demonstrations and benchmarks.
A star marks a capability described in Protocolis’s pending US provisional patent applications. Pending applications are not issued patents.
Open-source Agent, with optional Browser Use Cloud. Current source review; the benchmarks below retain their separately named versions.
Agent source · Cloud capabilities
Other dashed rows: not established in the sources linked above.
The documented OpenAI browser-tool integration with Browser Use as browser driver; this is a separate tool loop from Browser Use Agent.
OpenAI integration · Cloud capabilities
Other dashed rows: not established in the sources linked above.
Browser Use’s browser-harness CDP tool and skill. Reviewed independently of Browser Use Agent and the OpenAI integration.
README · Page workflow · Optional cloud service
Other dashed rows: not established in the sources linked above.
Current v4 documentation, including optional Browserbase services. The Stagehand demonstration uses its explicitly named older version; these are not new benchmark results.
Actions and variables · Observe · Extract · Caching · Browserbase CAPTCHA handling · Text-based result-check example · v4 extract screenshot default
Other dashed rows: not established in the sources linked above.
Microsoft’s Playwright MCP server, including its optional verification tools. Generic CDP connectivity is not treated as a built-in bot solver.
Other dashed rows: not established in the sources linked above.
Published contract for the hosted, closed-source service. The public repository is its client and documentation. The vendor marks some acceptance checks open; this review does not independently validate the service.
How it works · Current shipped scope · Vault and status
Other dashed rows: not established in the sources linked above.
keon/browser-control, the Rust CDP CLI. This is a distinct project inspired by Browser Harness, not a renamed Browser Use agent.
README · Final-answer checks · Optional Browser Use Cloud
Other dashed rows: not established in the sources linked above.
Protocolis is the work of a sole inventor. It is operational, patent pending and in use by its first paying customers, two trucking dispatch offices. The engine, the credential broker, the patent filings, the measurement record and the reproduction kit are offered together, with a transition period and, if the acquirer wants one, a continuing role.
A recorded demonstration and a live screen share.
Your engineers run their own agents and tasks on a hosted instance: the evaluation platform.
At letter of intent, a time-limited build for inspection.