A desktop SEO browser with an AI agent — a Chromium-based browser with a docked SEO analysis sidebar, a local change-history store, and an agent that can drive the browser and run SEO skills.
-
Phase 0 — Foundations: done. A pnpm/Turbo monorepo, the hardened Electron shell, the typed sender-guarded IPC contract, the security policy module, CI security gates, and macOS arm64 packaging.
-
Phase 1 — Browser core: done. A usable, hardened browser: tabs, an address bar, navigation guards for browsed content, per-origin permission prompts, page commands and a context menu, and a local browser-session database that restores tabs. The plan and how it was built:
docs/plans/phase-1-browser-core.md. -
Phase 2 — SEO engine and sidebar: done. Visiting a page gives a live audit with a 0–100 health score in a docked, collapsible sidebar: Overview, On-Page, Links, Images, Social, Schema, Indexability, Page Speed and Compliance panels, from 101 rules of frost's own. Extraction runs in isolated worlds, and the analysis runs in a worker thread with a hard timeout. The plan and how it was built:
docs/plans/phase-2-seo-engine.md. -
Phase 2.5 — SEO tooling depth: built (Tasks 1–7); its macOS release check is run separately. The day-to-day toolkit around the audit: per-tab render profiles (device, crawler user agents, network, JavaScript and cache off), a served-HTML vs rendered-DOM comparison, a Content panel with a snippet editor, a command palette, an SEO devtools drawer, find in page, reopen closed tab, history and a choice of search engine, and on-request link, sitemap and hreflang checks; the sidebar gains Content and JavaScript panels (eleven, with 111 rules). What the app does now:
apps/desktop/README.md; the plan and how it was built:docs/plans/phase-2.5-seo-tooling.md. -
Phase 3 — SERP, intent and indexability: built (Tasks 1–7); its macOS release check runs on the macOS CI runner. Two new sidebar panels (thirteen in all):
- SERP, on a Google, Bing or DuckDuckGo result page you load. It shows the page's composition, page-1 ownership by domain, outlines of the features on the page, the tab's search parameters, and an honest "couldn't read this layout" state. It reads the page with selectors that are signed data, updatable without an app release.
- Visibility, which on request fetches the page as Googlebot, Bingbot and AI crawlers fetch it, beside a normal browser, and marks the differences.
An optional on-device intent classifier labels queries with its confidence; its model is downloaded only after you agree. The plan and how it was built:
docs/plans/phase-3-serp-intent.md. -
Phase 4 — Accounts, entitlements, telemetry and Search Console: built. Sign in to a frost account through the system browser, confirmed in the app; a signed snapshot arrives that main verifies, with offline grace. frost is free and open source (2026-10-29): only sync needs an account; the AI agent and Search Console work signed out. Telemetry (usage traces and error reports) is off by default, scrubbed, and does what its two switches say. A read-only Search Console connection feeds a page and site performance panel, and a Settings view holds all of it. The account API is a lean backend (
apps/backend) that runs locally and in tests only. The plan and how it was built:docs/plans/phase-4-accounts-telemetry.md; the backend:apps/backend/README.md. -
Phase 5 — AI agent: built. A sidebar agent that audits and operates the browser, in a utility process of its own that holds no keys: main makes every model request (your own Anthropic or OpenAI key, the Vercel AI Gateway, or your ChatGPT plan; keys stay in the OS keychain). Read-only tools run on a standing grant; navigate, click and visibility probes ask unless granted for the site;
execute_javascript, AI search and MCP tools ask every time, and tool output reaches the model as untrusted data. SERP, intent, probes and Search Console come from Phases 3 and 4, and the agent needs no account (2026-10-29: only sync is gated). Skills, a grants page, chat retention and export, a context meter and MCP connections (labelled untrusted, sandboxed where the OS allows) round it out. The plan and how it was built:docs/plans/phase-5-ai-agent.md; the ChatGPT plan route:docs/plans/agent-chatgpt-plan.md. -
Phase 6 — Change intelligence: built (Tasks 1–8). Per-page change history and visual evidence, kept locally, with explicit retention controls:
- Changes (a sidebar panel) captures what a page says to search engines. Each distinct state becomes an immutable version, recorded with its render profile and search parameters, and a readable diff shows what changed, by region: head, HTTP, structured data, main content, navigation, header, sidebar, footer.
- Screenshots of the viewport, a selected area or the whole page (stitched as it scrolls), stored by content hash in the profile directory.
- Projects group pages and searches. Capture now visits each one in a tab of its own.
- Retention: keep the last N versions or N days per project, Purge now with a preview, measured disk usage, and a purge log.
- Inspector (a sidebar panel): a read-only DOM tree and element picker.
The database refuses changes to captured rows; they are removed only by a recorded purge. The plan, its schema and how it was built:
docs/plans/phase-6-change-intelligence.md. -
Phase 7 — Cross-platform packaging, signed auto-update and sync groundwork: built. Linux (AppImage, deb, zip) and Windows (Squirrel installer, zip) packages beside the macOS one, each verified in CI; an Ed25519-signed, staged update manifest, with nothing downloaded or installed without a click (builds are unsigned, so updates install by hand); and a libSQL driver seam with the groundwork for opt-in sync of change intelligence. The plan and how it was built:
docs/plans/phase-7-sync-update-platforms.md. -
Phase 8 — Free and AI-complete: built. 14 built-in SEO skills in four groups, a skills library you can search, switch off, export and import (with a review step), Gemini and Grok models on the AI Gateway route, a Reasoning setting, a catalogue of MCP server presets, and deeper readability (sentence length, difficult sentences, complex words; Spanish, French and German). The plan and how it was built:
docs/plans/phase-8-ai-complete.md. -
Phase 9 — Search Console depth and AI readiness: built. Any date range with country and device filters, URL Inspection, target-query suggestions from Search Console, an AI readiness score (assistants, their crawlers and robots.txt verdicts, pre-render coverage,
llms.txt), an on-device page type, and opt-in AI checks on your own model route (off by default). The plan and how it was built:docs/plans/phase-9-gsc-ai-readiness.md. -
Phase 10 — Projects as a workspace: built. Local rank tracking from the result pages frost captures (position, change, trend), comparison of two searches or two dates, bulk-added searches with tags, a project's site, competitors and favorites, CSV export, a SERP parameter editor, and the same search opened in Perplexity, ChatGPT or Google AI Mode. "Check rankings" and the capture cadence stay switched off (
RANK_CHECK_AVAILABLE) until their terms note is written. The plan and how it was built:docs/plans/phase-10-projects-rank-tracking.md. -
Phase 11 — Practitioner tooling: built (Tasks 1–13). The everyday tools around the audit:
- Copy and export: 24 page values, Images, Headings and Speed tables as CSV, and a Copy SEO Data context menu, all written by main from the audit; Copy JSON-LD, the page's or a suggested one.
- Highlights of internal or external links, lazy images, a keyphrase, or everything at once, and a fuller result preview (site name, schema breadcrumb, date, rating, rich cards).
- Devtools: the network log exported as HAR or CSV (no bodies; known sensitive URL parameters redacted, best effort, so review the file before sharing), and local and session storage keys. The Inspector searches the DOM and shows the rendered and served source.
- This tab only: CSS off, cookies off, or a browser location, beside the render profile.
- Screenshot editor: arrows, boxes, text and crop, saved as a copy.
- Agent: search, rename and archive chats; attach up to three other tabs to a chat.
- Workspace: resizable sidebar and agent panes, panels pinned to the rail, history retention and clearing by time range.
- Help: keyboard shortcuts, What's New, a setup checklist, and Export Diagnostics, a scrubbed report you review and save yourself (frost never sends it).
The plan and how it was built:
docs/plans/phase-11-practitioner-tooling.md. -
Next: finishing sync (the switch to turn it on, a worker thread, a real server), and code signing for macOS and Windows.
See search results as they look from another place, through your own proxy (Settings › Locations; "New tab in location…" in the command palette). frost bundles no VPN and recommends no proxy provider: you add an HTTP, HTTPS or SOCKS5 proxy you already use. Each location has its own temporary browsing session, kept in memory, with WebRTC limited to proxied traffic; its tabs carry a badge, are not kept in history and are not restored after a restart. A proxy's username and password stay in frost's main process, protected by your OS keychain; an HTTP proxy receives them unencrypted, so use an HTTPS proxy when it needs a sign-in. Location tabs never get your device location, and their permission answers last only until frost quits. What is not proxied: pages on localhost, loopback and link-local addresses are reached directly, and frost's own requests for a location tab (robots.txt, the served HTML, probes, link, sitemap and hreflang checks, a site crawl) and "Open in" tools still go from your own connection. Result pages read in a location tab record the location, so their rankings form a series of their own, and they stay on this device (they are never synced).
These are built, but they ship turned off until the project's review of their terms is done; until then the app shows them disabled, with a note saying so. Settings › Data sources takes your own keys for SerpApi, DataForSEO or Serper (result pages) and for Perplexity or OpenAI web search (AI answers). With them, a project's searches can be checked through the API ("Check via API"), and the AI answers view keeps what those engines answered, with their citations. The site crawl fetches a site's served HTML within robots.txt, paced, and audits every page. Nothing runs on its own: each request is one you start, with your own key (the crawl needs none), and what comes back stays on this device.
frost is a clean-room build. It is designed from a written behavioral/design
spec authored by us, not from any third party's source. This repository contains
no third-party code, bundles, or UI assets. Third-party open-source libraries are used
only as dependencies under their own licenses (see apps/desktop/THIRD_PARTY_NOTICES).
apps/desktop/— the Electron app (main, preload, the React chrome, the on-device intent utility process), packaged with Electron Forge.apps/agent-runtime/— the agent runtime (@app/agent-runtime): the chat loop, run in an Electron utility process with no network globals and no credential.apps/backend/— the account API (@app/backend): web and desktop sign-in, signed entitlement snapshots and an admin CLI. Hono, better-auth and Drizzle on Postgres (PGlite in dev and tests). Never deployed from this repo.packages/seo-core/— the SEO audit engine (@app/seo-core): the rule catalog's checks, the health score, the schema, robots.txt, hreflang, page speed and security readers, the content report, the served-vs-rendered comparison and the link, sitemap and hreflang check readers. Pure TypeScript, with no Electron, DOM or Zod dependency.packages/ai-skills/— the agent's tool definitions, their input schemas, the read-only data ports and the declarative SEO skills (@app/ai-skills).packages/change-intel/— the change-intelligence diff engine (@app/change-intel): canonical JSON, snapshot identity, the region-tagged diff between two captures and word-level inline diffs. Pure TypeScript, likeseo-core.packages/db/— the local database (Drizzle overnode:sqlite): the browser session, history, pinned render profiles, content notes and app settings, and change intelligence's append-only captures, versions, diffs, screenshots, projects and purge log.packages/ipc-contract/— the typed tRPC IPC contract and its sender guard.packages/telemetry/— consent-gated telemetry (@app/telemetry): OpenTelemetry traces and Sentry error events behind a consent gate, with allowlists and scrubbing. Pure Node, no Electron import.specs/— machine-checkable spec artifacts (constants, Zod schemas,*.feature.mdacceptance lists); every scenario maps to a named test.docs/design/IMPLEMENTATION-PLAN.md— the design doc (intent, stack, security model, phased roadmap).docs/plans/— the executable, spec-driven phase plans (Phases 0–12 and the ChatGPT plan route)..github/— the CI pipeline (workflows/ci.yml, one staged run ending in theCI passedcheck;specs/ci-pipeline.feature.md) and its shared install steps (actions/setup).semgrep/,security/,scripts/— the CI security gates' rules, baseline and runner,scripts/ci-plan.mjs(what a CI run covers, and its verdict), andscripts/sign-selectors.mjs(pnpm sign:selectors, for maintainers signing a selector pack) andscripts/sign-update-manifest.mjs(pnpm sign:update-manifest, for signing an update manifest).
- Node ≥ 24 and pnpm 12.15.0 (
package.json#packageManager);pnpm install. pnpm lint,pnpm test;pnpm --filter @app/desktop build && pnpm --filter @app/desktop test:e2e(needs a display);pnpm --filter @app/desktop startfor a dev run. More, including macOS packaging, inapps/desktop/README.md.- New work follows
CONTRIBUTING.md: spec first, a failing named test, then the code.
TypeScript + React + Electron Forge · pnpm/Turbo monorepo · typed tRPC IPC with a central sender guard · SQLite + Drizzle (local-first) · Vercel AI-SDK agent · security & consented telemetry designed in from commit #1.
Copyright (C) 2026 SnowOpsHQ.
frost is free software: you can redistribute it and/or modify it under the terms of the
GNU Affero General Public License, version 3 (AGPL-3.0-only), as published by the Free
Software Foundation. See LICENSE. It is distributed WITHOUT ANY WARRANTY; without even
the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
Bundled third-party libraries keep their own licenses; the packaged app ships their notices in
THIRD_PARTY_NOTICES. For a commercial license, contact SnowOpsHQ.