Web execution just entered the solid-state era.
The just-in-time compiler for agents.

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

Speed without the trade.

Protocolis AdaptiveRecommended
An automatic transmission for browsing agents.Faster than every product we’ve tested. Even cheaper than Standard. Highest measured completion.The default. The same engine as Standard decides, step by step and from its own evidence, how much model each step needs: a small choice model takes the routine steps, and the agent's model takes the rest.
Protocolis Standard
The fastest and cheapest full-model agent mode we’ve ever measured. Highest measured completion. Only Protocolis Adaptive costs less.Nothing between your agent and the page but compiled operations. Standard lets you evaluate the engine without Adaptive’s helper models. The same engine as Adaptive; the one difference is that your agent’s model makes every pick, with no choice model and no second agent. It calls each operation, reads each result, and completes as many tasks or more than every product we have measured it against.
One engine, two modes

The trade the others make

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.

one engine

One engine. Any model. Any task length.

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.

More done, for lesstasks completed · cost per completed task · time as depth
Protocolis AdaptiveProtocolis Standardring: driven by an outside agent (tools)every other system · point at a markSmallStrongFrontier modelfaint line back: time per completed task, further back is slower
Show as a table

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.

Head to head.

Same coding agent, model and prompt; fresh browsers on one Mac. Answers checked against site data or Browser Use’s judge.

54%more tasks completed
3.4×faster per completed task
2.2×cheaper per completed task
100%tasks completed vs Jev’s 65%

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

Detailed benchmarks
02

Every leading agent stack. One chart.

Loading benchmark records…
System strengthstasks completed · every model size and task length

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.

how it works
Your agentAny model, the one it already uses. It plans the task, calls operations by name and sees every result.
ProtocolisPatent pending
CompileReads the live page and compiles it into typed operations. No model. First contact, any site.
OperatePerforms the real clicks and keys for each operation, and waits for the page to settle.
VerifyReads the page after every act and returns what changed, or a refusal that names why. Checks that hold are kept; checks the evidence stops supporting are dropped.
The websiteAny live site, as it is. No selectors, no adapters, no taught workflows.
03compare

Feature for feature, it isn't close.

Sixteen capabilities, compared product by product.

CapabilityProtocolisBrowser UseOwn agentBrowser UseOpenAI toolsBrowser HarnessBrowser UseStagehandv4 docsPlaywright MCPMicrosoftUnbrowseHosted serviceBrowser 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 capabilities16 / 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.

Comparison details & sources

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.

What each row means
★ Just-in-time cold compiler engine
Cold means compilation on first contact with no site-specific adapter, prior training or recorded workflow. A compiler that requires recorded workflows does not meet this criterion.
★ Automatic transmission: one mode that shifts step by step
One mode decides at every step whether the compiled engine can carry it or full model reasoning is needed, on first contact, so the caller never chooses between a fast mode and a full mode. Automatic reuse of earlier cached or learned results receives partial credit.
★ Self-healing: verification that learns from outcomes
Verification learned from the engine’s own record of outcomes: checks that confirmed are promoted, checks the live evidence stops supporting are demoted or withdrawn, and each is re-checked against the live page before use. Model re-inference or retries after a failure receive partial credit.
★ Application operations compiled on first contact
Derives callable application-specific operations from the live application on first contact, without a site-specific adapter, taught workflow or prior recording. Recorded request-chain compilation does not meet this first-contact criterion.
★ Compilation without model calls
The compilation step itself requires no model inference; model-free replay alone is not enough.
Screenshot-free execution and result checks
Full support requires a documented first-contact workflow that both executes actions and checks the resulting page state without screenshots, while retaining its documented verification path. Text selectors or switching vision off alone are insufficient. This concerns outcomes observable in page data, not visual-only judgments. Agent-directed checks are distinguished from automatic verification in the separate verification rows.
★ Action effects verified on first contact
Checks action effects against live application evidence on first contact, without site-specific training or a previously recorded workflow. Agent-authored checks and generic assertions receive partial credit.
★ Task completion verified on first contact
Checks the requested task outcome against live application evidence on first contact, without site-specific training or a previously recorded workflow. A learned operation’s declared outcome does not establish this first-contact capability. Model judges and caller-supplied assertions receive partial credit for checking the task, with the limitations stated below.
★ Compiled knowledge reused across visits
Persisted application-derived operations are reused on later visits. Cached model-generated actions receive partial credit; manually authored scripts alone do not.
★ Verified replay with protected-action stops
Replay checks current application evidence and stops before protected actions. History reruns, retries or partial policy gates receive partial credit.
Built-in bot-protection handling
The product includes handling for bot-protected sites. A separately supplied cloud browser receives partial credit. No rating promises success against every protection.
Challenge handling and isolated escalation
Handles simple challenges and escalates difficult challenges in an isolated flow. Solving or waiting without the full escalation mechanism receives partial credit.
Challenge lessons shared across sites
Challenge-handling lessons transfer between sites; a site’s earlier history does not prematurely terminate a new attempt.
★ Credentials bound to verified fields
Secret values stay out of model context and are released only to verified fields on the current live page. Placeholders, masking or domain scoping alone receive partial credit.
★ New API keys captured model-blind
Newly generated API keys are captured into protected custody without passing through model-visible output. Redaction alone is not capture.
Passkeys
A built-in broker supports passkey registration and authentication for agent workflows, beyond reusing an already signed-in browser.
Browser Use · own agent · 0 full, 8 partial

Open-source Agent, with optional Browser Use Cloud. Current source review; the benchmarks below retain their separately named versions.

Self-healing: verification that learns from outcomes — Related capability with limits
Failed steps are retried with the model up to a failure limit. Verification learned from earlier outcomes is not established.
Screenshot-free execution and result checks — Related capability with limits
Vision is enabled by default and the normal completion judge uses screenshots. Disabling vision and the visual judge permits text-only execution but removes that verification path, so it does not earn a full screenshot-free execution-and-checks rating.
Action effects verified on first contact — Related capability with limits
Browser-state feedback and individual action checks support agent verification; general application-effect proof is not enforced.
Task completion verified on first contact — Related capability with limits
The optional model judge evaluates completion. That is related functionality, not independent application proof.
Verified replay with protected-action stops — Related capability with limits
Saved history can be rerun with retries. Full live verification and protected-action stops are not established.
Built-in bot-protection handling — Related capability with limits
Stealth and CAPTCHA handling are supplied by optional Browser Use Cloud, rather than the open-source library.
Challenge handling and isolated escalation — Related capability with limits
Cloud CAPTCHA solving is documented; isolated escalation with the complete row contract is not established.
Credentials bound to verified fields — Related capability with limits
Sensitive-data placeholders can be scoped by domain. Verified live-field binding is not established.

Other dashed rows: not established in the sources linked above.

Browser Use · OpenAI tools · 0 full, 5 partial

The documented OpenAI browser-tool integration with Browser Use as browser driver; this is a separate tool loop from Browser Use Agent.

Screenshot-free execution and result checks — Related capability with limits
Text and accessibility reads support screenshot-free interaction, but the reviewed integration does not establish a complete first-contact result-check workflow without screenshots. It calls for independent verification and documents screenshot fallbacks.
Action effects verified on first contact — Related capability with limits
The caller can read page state to check effects. Action confirmation is an authorization callback, not an outcome verifier.
Task completion verified on first contact — Related capability with limits
The integration calls for independent verification of important external actions; no automatic task-proof mechanism is documented.
Verified replay with protected-action stops — Not established in reviewed sources
Read retries are documented, but clicks and typing are explicitly not replayed after uncertain effects.
Built-in bot-protection handling — Related capability with limits
The documented cloud setup can use Browser Use Cloud’s separate stealth and CAPTCHA service.
Challenge handling and isolated escalation — Related capability with limits
Cloud challenge solving is related functionality; the complete isolated-escalation contract is not established.
Credentials bound to verified fields — Not established in reviewed sources
The Agent’s sensitive_data mechanism is not established for this distinct integration, so its credential score is not inherited.

Other dashed rows: not established in the sources linked above.

Browser Harness · 1 full, 4 partial

Browser Use’s browser-harness CDP tool and skill. Reviewed independently of Browser Use Agent and the OpenAI integration.

Screenshot-free execution and result checks — Documented match
The documented workflow uses the accessibility tree to act, then targeted JavaScript or page_info to check the effect. These checks are agent-directed. Screenshots are still advised when layout or imagery is the task.
Action effects verified on first contact — Related capability with limits
The skill directs the agent to check effects with targeted JavaScript or page_info; enforcement remains with the agent.
Task completion verified on first contact — Related capability with limits
The workflow asks the agent to verify results; no automatic application-level completion proof is established.
Compiled knowledge reused across visits — Not established in reviewed sources
Reusable scripts and site skills are supported, but they are not automatically compiled application knowledge.
Built-in bot-protection handling — Related capability with limits
Optional Browser Use Cloud supplies managed stealth; this is a separate cloud service.
Challenge handling and isolated escalation — Related capability with limits
The optional cloud service supplies CAPTCHA solving. The complete isolated-escalation contract is not established.
Credentials bound to verified fields — Not established in reviewed sources
Existing browser sign-in and SSO are usable, but a model-blind, verified-field credential broker is not established.

Other dashed rows: not established in the sources linked above.

Stagehand · 1 full, 9 partial

Current v4 documentation, including optional Browserbase services. The Stagehand demonstration uses its explicitly named older version; these are not new benchmark results.

Automatic transmission: one mode that shifts step by step — Related capability with limits
On Browserbase, the server-side cache returns earlier act() results without model calls when the inputs repeat. The caller still chooses the call and the model; there is no step-by-step choice on first contact.
Self-healing: verification that learns from outcomes — Related capability with limits
With selfHeal on, Stagehand re-infers an action with the model when its recorded selector breaks and retries once. Verification learned from earlier outcomes is not established.
Screenshot-free execution and result checks — Documented match
The documented form workflow follows actions with an extract check of the login result; extract defaults to no screenshot. This is a caller-requested, model-based page-state check, not automatic application proof.
Action effects verified on first contact — Related capability with limits
Observe/extract and caller-authored assertions can check state, but do not impose automatic action-effect proof.
Task completion verified on first contact — Related capability with limits
The caller can assert completion from extracted page data. Automatic application-level task proof is not established.
Compiled knowledge reused across visits — Related capability with limits
Browserbase’s server cache reuses eligible model-generated act, observe and extract results. Initial inference is still required; local-browser caching differs.
Verified replay with protected-action stops — Related capability with limits
Observed actions can be reused and self-healing can retry failures. Full protected-action gating is not established.
Built-in bot-protection handling — Related capability with limits
Hosted Browserbase supplies stealth and CAPTCHA handling; this is an additional service.
Challenge handling and isolated escalation — Related capability with limits
Hosted CAPTCHA handling is documented; the full isolated-escalation contract is not established.
Credentials bound to verified fields — Related capability with limits
Variables can keep actual values out of model instructions. The documentation warns that server caching can receive values unless cache:false is used; verified-field brokerage is not established.

Other dashed rows: not established in the sources linked above.

Playwright MCP · 1 full, 3 partial

Microsoft’s Playwright MCP server, including its optional verification tools. Generic CDP connectivity is not treated as a built-in bot solver.

Screenshot-free execution and result checks — Documented match
Accessibility-based actions can be followed by text/value verification tools under the testing capability. The caller chooses assertions and expected values. Screenshots are not required for those page-state checks.
Action effects verified on first contact — Related capability with limits
Assertions and snapshots let the caller check effects; they do not automatically verify every application action.
Task completion verified on first contact — Related capability with limits
Verification tools can assert page state when the caller chooses. No automatic task-level proof requirement is established.
Credentials bound to verified fields — Related capability with limits
Secrets substitution and response redaction are documented conveniences, explicitly not a security boundary or verified-field broker.
New API keys captured model-blind — Not established in reviewed sources
Runtime URL-token redaction is documented, but that does not establish model-blind capture of newly generated API keys into a vault.

Other dashed rows: not established in the sources linked above.

Unbrowse · 2 full, 5 partial

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.

Automatic transmission: one mode that shifts step by step — Related capability with limits
Each request uses a learned capability’s HTTP replay when one is warm and falls back to rendering in a browser otherwise. This depends on a previously learned capability, not first contact.
Self-healing: verification that learns from outcomes — Related capability with limits
Tracks each learned capability’s health from real runs and marks failing ones; it falls back to browser rendering when replay cannot answer. Learned verification of each step is not established.
Just-in-time cold compiler engine — Not established in reviewed sources
Its documented compiler learns from recorded browser sessions or HAR files. That does not meet the cold criterion of compilation without a prior site-specific workflow recording.
Application operations compiled on first contact — Not established in reviewed sources
Documents callable HTTP operations learned from recorded sessions or HAR files. This does not establish application operations compiled on first contact without prior recordings.
Compilation without model calls — Not established in reviewed sources
The documented learn algorithm does not establish that every compilation path runs without model inference.
Screenshot-free execution and result checks — Not established in reviewed sources
Documents screenshot-free HTTP replay and declared-outcome checks for learned capabilities. The reviewed sources do not establish the same complete execution-and-result-check path on first contact without a learned workflow.
Action effects verified on first contact — Not established in reviewed sources
Packages declare guards and outcome checks for learned capabilities. The reviewed sources do not establish built-in checks of action effects on first contact without a previously recorded workflow.
Task completion verified on first contact — Not established in reviewed sources
Documents checks against a learned capability’s declared business outcome. The reviewed sources do not establish verification of the complete requested task on first contact without a learned workflow.
Compiled knowledge reused across visits — Documented match
Persisted learned capabilities are reused on subsequent requests. Public-registry reliability is separately marked as an open acceptance check.
Verified replay with protected-action stops — Related capability with limits
Verified outcomes, guards and no automatic retry of uncertain writes are documented; complete protected-action stops at every replay step are not established.
Built-in bot-protection handling — Documented match
The service includes browser-earned clearance, Chrome-fingerprint transport retries and browser/render fallbacks. This does not promise bypass of every challenge.
Challenge handling and isolated escalation — Related capability with limits
Waits out simple challenges and can request same-session verification. The full isolated-escalation contract is not established.
Credentials bound to verified fields — Related capability with limits
Documents model-blind vault references and autofill. The reviewed shipped-scope sources do not establish the full verified-live-field release contract; autofill acceptance checks are marked open.

Other dashed rows: not established in the sources linked above.

Browser Control · 1 full, 4 partial

keon/browser-control, the Rust CDP CLI. This is a distinct project inspired by Browser Harness, not a renamed Browser Use agent.

Screenshot-free execution and result checks — Documented match
The documented workflow uses snapshots and actions followed by text, eval or CDP evidence checks. Verification is agent-directed; this is not an automatic proof guarantee or a claim about visual-only tasks.
Action effects verified on first contact — Related capability with limits
The agent can inspect page state and check action results; automatic effect proof is not established.
Task completion verified on first contact — Related capability with limits
The bundled final-answer skill requires evidence and constraint checks. These are agent instructions, not a runtime completion verifier.
Compiled knowledge reused across visits — Not established in reviewed sources
Scripts and domain skills can be reused, but the tool does not document automatic application compilation.
Built-in bot-protection handling — Related capability with limits
The CLI explicitly supports provisioning Browser Use and other cloud providers. Their stealth is external to the CLI.
Challenge handling and isolated escalation — Related capability with limits
A separately provisioned Browser Use Cloud browser can provide CAPTCHA solving; the CLI does not establish isolated escalation.

Other dashed rows: not established in the sources linked above.

04acquisition

Offered for outright acquisition.

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.

Vahagn Boudakian, inventor.

vahagn@protocolis.co
Download overview 1 page · PDF ↓
  1. I
    Watch

    A recorded demonstration and a live screen share.

  2. II
    Try

    Your engineers run their own agents and tasks on a hosted instance: the evaluation platform.

  3. III
    Hold

    At letter of intent, a time-limited build for inspection.

Request access.

Tell me what you’d like to evaluate. Your request goes directly to Vahagn Boudakian.