Skip to content

Changelog

1.5.0 — 2026-11-19

  • Advertising, respected and delineated. Ads on pages (ad containers, rel="sponsored" links) are taken out of the content and relayed in a separate sponsored block, never learned and never given to the AI. Site owners can add agent ads (AIAgentAd, AIAgentAdsFile), targeted by path or keyword, with weighted rotation, AIAgentAdsRequired / AIAgentAdsTerms, impression and click logging (ads.log), and a click redirect limited to configured destinations. The AI is instructed never to put ads in answers.
  • MCP server at /.well-known/ai/mcp, so Claude, ChatGPT and other assistants can use any site as a connector. Tools: site_overview, open_page, ask, list_actions, do_action. A server-side cookie jar per Mcp-Session-Id keeps logins working across calls. Tested with the official MCP Python SDK.
  • Browser extension "AI Chat" for Chrome, Brave, Edge, Opera, Vivaldi, Firefox and Safari:
    • side panel or sidebar or popup, with forms built from field descriptions
    • a confirmation card, a labelled Sponsored box, and secrets masked in your own messages
    • shared login with the tab, and a chat that follows your browsing
    • tested in real Chrome; the Firefox build lints clean
  • The knowledge directory is now created whenever actions are on, so chat and MCP sessions work on sites without an AI service.
  • chat_core refactor shared by chat and MCP.

1.4.0 — 2026-11-19

  • Chat on any URL (ai-chat/1). Accept: application/vnd.ai-chat+json or ?ai=chat turns any page into a conversation. Each server turn says where the agent is, what it can do (you_can, with fields) and where it can go (you_can_go). The agent replies with say / do / go. The server keeps the location; the client keeps the cookies.
  • Every HTML page advertises its chat version: Link: <url>; rel="alternate"; type="application/vnd.ai-chat+json".
  • Browser-extension origins (chrome-extension://, moz-extension://, safari-web-extension://) are accepted, ready for the browser plugin.
  • /act refactored into a shared engine (act_core) used by both /act and chat.
  • Fixed: the cached action catalog could point to freed request memory after a cache refresh.
  • Added tools/chat_client.py (reference client), tests/mock_ai.py (mock AI provider), and docs/PROTOCOL.md, HANDOFF.md, LOVABLE_BRIEF.md.

1.3.0 — 2026-11-17

  • Actions: /actions catalog and /act. Agents do what a person does by clicking: forms opened fresh with the agent's own session, CSRF tokens, defaults, chosen button, redirects followed, cookies returned.
  • Functionality learned from public forms, PHP handlers ($_POST, required fields), Express/Laravel POST routes, and fetch/axios/$.post calls.
  • Gates: needs_login, needs_confirmation (payments, deletions, password changes), invalid_fields, read-only mode.
  • Security: CSRF hardening, the agent's real IP kept on internal requests, actions.log without field values, rate limit.
  • Fixed: a PHP guard false positive (the guard consequence must be inside the if body); PHP's no-store no longer blocks learning of anonymous pages.

1.2.0 — 2026-11-17

  • Learning modes: always / until-ready / off. Pages that disappear are forgotten.
  • Code scan (PHP, Laravel, Express, React, Vue/Angular, Next.js, .htaccess) to find which pages require login. Used for learning, guides, sitemap and the AI. Admin report in access.txt.

1.1.0 — 2026-11-17

  • AI service (Anthropic / OpenAI / OpenAI-compatible), configured once.
  • Learning gate: "I don't know this site well enough yet" until the learning period and page minimum are met. ready file override, immediate activation.
  • Conversations, answer cache, rate limit, fallback on AI failure, tools/prime.sh.

1.0.0 — 2026-11-17

  • First version: agent detection, exact / guide / text modes, English sitemap, /ask keyword answers, /llms.txt, access-checked crawl.