Everything it does.
Nothing it pretends to.
Every feature on this page ships in the current build unless it's tagged otherwise. Preview means it works but is still rough; roadmap means it doesn't exist yet.
Ask — answers with receipts.
Point Spike Local at your folders once. From then on, any question — "what did we quote them in March?", "ใครรับผิดชอบเฟสสอง?" — gets answered from the actual documents, with numbered citations linking every claim to its source.
- Semantic search over Word, PDF, Excel, text and more — tables included
- Scope answers to an area: one customer, one project, one case
- Cross-customer contamination guards — customer A's question never quotes customer B's files
- Thai and English retrieved together, natively
Data migration moved from May 12 to June 2 after the schema change 1, and UAT start shifted two weeks with it 12. Go-live is unchanged 2.
Documents — drafts grounded in your brain.
"Write the proposal for the upgrade project" becomes a real, structured document — every figure and claim pulled from your own files and cited, never invented. Long documents build chapter by chapter, and each chapter passes a fact-check before you read it.
- Chaptered builder for long documents, with per-chapter quality checks
- Unverified sentences flagged inline, with a one-click fix menu
- Targeted revision — "rewrite chapter 3 shorter" touches only chapter 3
- Optional per-chapter web research when you want outside context
- Clean PDF export with headers and footers
Capture — meetings in, minutes out.
Record a meeting or drop in any audio or video file. Spike transcribes it on your machine — with a Thai-tuned speech model when the room speaks Thai — then writes structured minutes: summary, decisions, action items.
- Local transcription — recordings never leave your disk
- Staged pipeline picks the right model and recovers from failures itself
- Minutes, clean-up, and summaries in the transcript's own language
- Screen recordings get described too — what was on screen, when
Brain — a memory you can argue with.
Overnight, Spike reads your indexed files and extracts short, verbatim facts — deadlines, values, owners, commitments — each pinned to its source. The Facts screen is where you verify what it believes: confirm, correct, or delete anything.
- Every fact is a quote from a real file, never a paraphrase
- Contradictions between documents surfaced automatically
- Organize everything into areas — customers, projects, products
- Facts power answers only for the right subject, gated by entity
Models — yours to choose, place, and replace.
Spike ships with sensible defaults and lets you go as deep as you like: pull any Ollama or Hugging Face GGUF model, assign models per job, and route all local inference to a stronger machine when you have one.
- Per-job roles — chat, deep answers, Thai, vision each pick their model
- Remote model host: your GPU box over LAN, an authenticated tunnel, or Ollama Cloud
- Honest badges — every answer says Local, LAN, Web, or Cloud
- Hardware-aware defaults — a GPU machine gets the quality model automatically
Workspaces — real analysis, not vibes.
Drop a spreadsheet into a workspace and run deterministic analyses — the numbers come from code, and the model only writes the narrative around them. Vertical packs add domain playbooks.
- Financial analysis, trends, and forecasting run as real computation
- Finance, insurance, and medical packs with ready-made deliverables
- Results become cited, exportable documents
Research & Autopilot — the rough edge.
Deep research plans first, then works through sources and writes up findings — solid today on a capable model. Desktop autopilot (Spike operating your apps by looking at the screen) is real, gated behind explicit permission, and honestly still early. Both are labeled preview because that's what they are.
- Research: plan-first deep dives with a quick or thorough mode
- Browser tasks: Spike can open, read, and fill pages in a side panel
- Desktop autopilot: three independent permission gates, a visible stop button
The quiet features that make it livable.
Carry your whole brain between machines in one file — you choose what it contains. Secrets are never included.
Add any provider with your own API key. Trust zones and redaction decide what it may ever see.
Every long-running job — answers, drafts, indexing — has a stop button that actually halts the model.
Drop in screenshots and photos — Spike reads figures, tables, and text from images on-device.
Teach Spike repeatable jobs and keep a task list it can act on — all stored locally.
macOS and Linux builds, encrypted multi-device sync, and team vaults are built or in progress — not shipping yet. The changelog tracks what's real.
See it on your own files.
The demo that matters is the one where it answers a question only your documents could answer.