Changelog
What's new
Write mode and downloads
- Write mode: an optional editor for long documents. Try alternative words, sentences and paragraphs in place and flip between them with the arrow keys; fade text without deleting it; keep an overflow notepad next to your draft.
- The Lab: AI marks typos, weak or long sentences, off-tone words and filler, and suggests trims of 10 to 50% by fading text. It never rewrites your words. Uses your own AI key.
- Drafts save on their own; Publish updates the page. Markdown stays the default editor and the published format.
- Download any page as PDF, Word (.docx) or Markdown.
- Security hardening: owner roles, single-use invites, rate limits, versioned sessions and page history with restore.
Board, Claude connector and EU AI Act transparency
- Board: a kanban where cards go to people or agents. Agents pick up their cards over MCP, comment, and hand back for review.
- Claude connector: add Groundtruth as a custom connector in Claude and sign in with OAuth. Invite links for teammates.
- EU AI Act Art. 50: AI-written answers and AI-extracted contract terms are labelled and carry a signed, machine-readable provenance mark. Anyone can verify it at /api/provenance/verify with the public key at /.well-known/ai-provenance.
- Agents are marked as AI everywhere they appear, and pages last edited by an agent say so, in the app and in the markdown twin.
- New AI transparency page at /ai.
Agent Store + bring your own key
- Agent Store with three agents: Contract Agent (cancel-by dates and deadline alerts, AI contract extraction), Churn Agent (explained churn-risk scores from customer interactions) and Freshness Agent (stale-page alerts).
- Alerts inbox, Slack-compatible webhook notifications, and daily scheduled runs.
- New MCP tools: list_alerts, resolve_alert, list_agents, run_agent, log_interaction.
- Bring your own Anthropic or OpenAI key for answers in the app. Keys are encrypted, test without spending tokens, and are never shown again.
- Agents asking questions now get the cited sources and compose answers with their own model, at no token cost to anyone but them.
Groundtruth public beta
- Workspaces with pages, decisions, records and an attributed activity log.
- Ask with citations — Claude-written answers grounded only in your workspace, or extractive answers when no model key is configured.
- MCP server (Streamable HTTP) with 14 tools.
- REST API mirroring every MCP tool at /api/v1/<tool>.
- Markdown twins for every page (/p/<id>.md) and a per-workspace llms.txt.
- Staleness: pages not verified in 90 days are flagged to humans and agents.
- Retrieval tuned with an autoresearch loop: +0.23 eval score in-sample, +0.06 on a held-out set, zero false answers on unanswerable in-sample questions.