
- Replay calls no model: a session recorded once compiles to semantic locators with pre- and post-conditions
- Repairs are reviewable diffs, grouped by root cause, tested against the page they came from and reversible
- 34 agent tools over one executor, shared by the sidebar, the terminal and any MCP client
Most browser automation should not call a model. FullSpectrum records a session once and compiles it into an artifact of semantic locators with pre- and post-conditions per step, then replays it deterministically against a pinned Chrome. When you do want the browser to act for you, a governed agent does, planning and acting on the page in front of you and writing every step to a signed audit trail you can budget and review.
The Problem: Automation That Guesses
Browser agents re-think every run, and record-and-replay tools heal silently.
- A model-driven run costs tokens every time and can do something different each time
- When a locator drifts, incumbent tools substitute an element and move on, and nobody can review what changed
- A credential typed during a recording ends up inside the artifact
- An agent driving a browser nobody can see is a keylogger with manners
- Nobody can prove afterwards which page's content was on screen when an action was taken
The Solution: A Runtime That Thinks Once
FullSpectrum separates the two jobs. Workflows are compiled artifacts that replay without a model, stop when a page drifts and propose a repair as a reviewable diff. The live agent works in the browser on screen, through the same tools and the same executor, with the user's policy asked before anything mutating or sensitive. Both write to a hash-chained, Ed25519-signed audit log with per-step page provenance.
Key Features
Record once, replay forever
A headful recording turns clicks, typing and choices into steps whose semantic identity is computed in the page at the moment of the click. Replay is statically validated before it touches the browser: whatever a step requires, the step before it must guarantee. Passwords are never recorded; a step carries a secret name and replay refuses to start without it.
Authoring without re-recording
Delete, move, insert and parameterise steps, re-record from the middle, assemble workflows from pieces of others, clone a flow onto another host and dry-run it against a page without acting. Every save is numbered, diffable and reversible.
Healing you can read, diff and approve
Relocation uses similarity rather than a model, healing is off by default, and a repair is explained in business terms, tested against the accessibility tree it came from, grouped so one page change is one decision, and rolled back as a forward inverse edit. A canary finds drift before a business run meets it, and alerts reach Slack, Teams or PagerDuty without ever failing the run they report on.
A governed live agent
Thirty-four tools over one executor, shared by the in-browser sidebar, the terminal and any MCP client, so what the agent can do cannot drift between them. Targets are refs resolved against a fresh snapshot, an ambiguous click is refused rather than guessed, and every action lands in the window on screen with its origin named.
Identity, audit and extraction
Profiles with their own vault, sessions and cookie jar, keys sealed behind a passphrase or the OS keychain, an append-only audit log with per-step provenance, and typed extraction that gets the data out through a REST API with TypeScript and Python SDKs.
Pages and apps as JSON
A page is an authored semantic document rendered to clean HTML, a live page can be captured and mirrored into one, and a single file describes a whole application with identity, routes and data.
Strategic Vision
FullSpectrum is the browser layer of the AW3 ecosystem, built on the same approval-first principle as TRMNLY: an agent that reaches the whole machine and runs nothing without consent. It exposes the runtime to Claude Code and any other MCP client, and its workflows are files that live next to the code they test.
Its foundation enables:
- Browser automation priced by the run rather than the token
- Repairs a reviewer can approve with evidence instead of imagination
- A live agent that acts where the user can see it, under the user's policy
- An audit trail that answers the question an incident actually asks
Status and Stack
Status: live at fullspectrum.one with the Local tier free forever; hosted Pro and Team tiers on a waitlist while the hosted store and fleet-managed Chrome are built
Platform: native macOS browser, with an iOS reader and an Android browser in v0 and Windows and Linux in scope
Stack: a TypeScript monorepo of 19 packages over the Chrome DevTools Protocol with a pinned Chrome for Testing; an AppKit shell over CEF Chromium with a Rust sidecar; SwiftUI on WKWebView for iOS and Java over the system WebView for Android; TypeScript and Python SDKs, an MCP server and a hand-written static site; published under the PolyForm Noncommercial licence


