Transactional write safety
Preconditions, idempotency keys and a ledger that separates verified from indeterminate.
Up to three independent native WebKit pages for your MCP tasks, sharing your saved sign-ins, with explicit native approvals and observable action results. Developer Preview 0.6.42 for supported Macs.
The browser runs on your Mac. Observations go to your MCP client and potentially its model provider. External Safari login cannot resume this MCP session.
Most automation reports the first and hopes for the other two. WebKitUI MCP treats them as separate facts, each with its own proof.
A JavaScript call fired an event. The page may have seen it — or ignored it as synthetic.
Native input can produce trusted browser events. This does not prove that a person approved the action.
A postcondition verifies the observed page state. A server-side commit needs independent backend evidence.
Four states most automation stacks conflate, kept strictly separate: what exactly was requested, who authorized it, how it was dispatched, and what proves it happened.
Every write binds to one exact target, arguments and idempotency key, with a short expiry. Change anything, and the binding is void.
A human approves that exact binding. Approval is a decision, not an action — the model cannot mint or forge it.
Only then does native AppKit actuation run the write, in a real WKWebView, against a freshly re-resolved semantic target.
Exact page-state postconditions are checked and a durable receipt recorded. Backend persistence requires separate evidence.
An indeterminate action is reconciled by re-observation of its postcondition. It is never blindly replayed.
Those are different, legitimate jobs — with better tools for them. See the comparison below.
| Dimension | WebKitUI MCP | Playwright MCP · Selenium | Managed cloud browsers |
|---|---|---|---|
| Local human authority | Built in — exact per-write approval on your Mac | Scriptable; typically runs unattended | Remote, policy-based controls |
| Input mechanism | Native AppKit actuation in a real WKWebView | Browser automation protocols; event behavior depends on the action | Depends on provider and browser backend |
| Direct state verification | Postconditions + durable receipts, by default | Assertions you write and maintain yourself | Varies by provider |
| Scale & concurrency | Deliberately low — bounded sessions per client | High — parallel contexts, grids, CI | Very high — elastic fleets |
| Ecosystem breadth | Focused: macOS, WKWebView, MCP | Very broad: every major engine and language | Broad APIs and integrations |
| Managed infrastructure | None needed — runs on your Mac | Self-managed or CI-hosted | Fully managed by the provider |
Pick Playwright, Selenium or a cloud fleet for breadth, scale and CI coverage. Pick WebKitUI MCP when a small number of writes must be approved, trusted and proven on a Mac you control. Many teams will use both.
Ordinary navigation and approved clicks don't ask for Touch ID — you'd never accept the friction, and they don't need it.
Releasing a credential is different. SiliconPass requires fresh device-owner authentication at the moment of release, using whichever method your Mac has available.
Availability is checked locally at the moment of release and is never inferred from lid state. If macOS offers no method, no secret is released and there is no automatic retry.
Fresh at the moment of release — never cached approval, never a stored token handed to the model.
Commercial access on request. Uses outside the BUSL grant need a separate agreement.
Contact LorisLabs for an agreement. This preview does not require an activation key; the licence defines permitted use.
BUSL-1.1 governs the preview. Commercial rights require a separate accepted agreement.
The public repository carries the architecture docs, the test suite and the benchmarks behind each of these categories — reproducible on your own machine.
Preconditions, idempotency keys and a ledger that separates verified from indeterminate.
AppKit-dispatched gestures in a real WKWebView, measured rather than assumed.
URL, checked, selected, enabled, attribute, value, dialog, option or semantic evidence — never “the UI looked right”.
Indeterminate writes re-observed until proven or abandoned — never blindly replayed.
SiliconPass release under fresh device-owner authentication; secrets never reach the MCP client.
Site-by-site behavior recorded as dated evidence, revalidated per version — not marketed as universal.
Run it against your own consoles for a month. If the receipts earn their keep, license it.