Shipping in public.

Run your models anywhere — honestly labeled.

The remote model host arrived: point every local-model feature at a stronger machine, and know exactly where each answer ran.

• Route inference to a GPU box over your LAN, through an authenticated tunnel (API key, custom header, or basic auth), or via Ollama Cloud
• Answers served remotely carry a LAN badge; the privacy report names the real host
• Unreachable host = a clear error and one-click "use local for now" — never a silent fallback
• Automatic fact-check pass on every chat answer, against its own sources

Long documents that check themselves.

The chaptered document builder graduated: real structure, real verification, and honest routing under the hood.

• Chapter-by-chapter drafting with per-chapter fact-check before you read it
• Unverified sentences flagged inline with a one-click fix menu
• Optional per-chapter web research for outside context
• Semantic intent routing replaced a pile of brittle keyword rules — fewer misfired requests
• Desktop autopilot shipped as a gated preview: explicit permissions, visible stop

A brain you can carry — and question.

• Backup & Restore with contents you choose; secrets never ride along
• Facts screen redesigned around verification: conflicts first, confirm-on-use
• Knowledge-graph answers gated by subject — customer A's question can't quote customer B
• Multi-file chat attachments with per-file citation attribution
• Hardware-aware model routing — GPU machines get the quality model by default

The foundation line.

• Ask: cited answers across Word, PDF, Excel and more — Thai and English together
• Capture: local transcription with a Thai-tuned speech model, minutes and summaries
• Spike Doc: grounded drafting with citations and PDF export
• Areas: organize the index by customer, project, or case
• Workspace analyses: deterministic financial computation with vertical packs
• Trust zones, redaction, and the in-app privacy report