Korvo

Korvo is a local-first desktop application built by Snab Limited (trading as Korvo, Company No. 16006744, England & Wales). It runs on macOS, with Windows and Linux in alpha, and gives knowledge workers a private workspace for turning scattered documents into traceable, cited decisions. No account is required to download and start using the free Explorer tier.

What It Is

Korvo positions itself as a "decision engine" rather than a chatbot or generic AI workspace. Users drop in PDFs, spreadsheets, notes, and transcripts; Korvo indexes them locally using its on-device embedding engine (Korvo Embed), plans the analysis, executes grounded synthesis with inline citations, and preserves the full decision record — including the plan, evidence, model details, revisions, and unresolved questions — for later inspection. The core workflow is: Capture → Structure → Plan → Execute → Preserve.

Local-First Architecture and Privacy Model

Korvo's privacy design is structural rather than policy-only:

• Files, notes, the local SQLite index, plans, and decision history reside strictly on the user's device. Korvo states it never holds workspace data on its servers.
• BYOK (Bring Your Own Key): Users connect OpenAI, Anthropic, or Google directly with their own API keys. Prompts and selected context go to that provider under the user's own agreement; Korvo does not proxy, inspect, or log them.
• 100% on-device option: Connect Ollama for full offline operation — no prompt, project context, or document data leaves the machine.
• Korvo Embed, the on-device embedding engine, is a 34 MB encoder with no HTTP client in its dependency tree, making it structurally incapable of sending documents to a network. It is verifiable in one command and enforced in CI.

Key Capabilities

• Plan Mode: Korvo detects the goal, surfaces assumptions, flags missing context, and proposes analysis steps before any generation runs. Users approve the plan.
• Grounded outputs with citations: Every material claim links back to the specific source file, sheet, and cell range in the project.
• Two-model critique pass (beta): One model builds the case; a second critiques it. The revision and unresolved questions remain visible.
• Decision history: Conclusions, source links, model details, timestamps, and open questions are preserved together and exportable to PDF or Markdown.
• MCP integration: Users can connect the built-in Superhuman Mail preset or a custom MCP server. Chat models cannot autonomously execute MCP tools; mutating or unannotated tools require explicit confirmation.
• Medha: An on-device quality-measurement layer that tests retrieval with deterministic metrics (recall, MRR, nDCG) in under 10 ms per operation, with zero network calls.

Target Audience

Korvo is built for professionals who need to show their work:

• Investors running due diligence across decks, financials, founder notes, and call transcripts.
• Founders turning board material, customer evidence, and strategy notes into plans with explicit assumptions.
• Operators synthesising multi-report context into actionable briefs.
• Researchers organising literature sets, comparing claims, and producing cited synthesis.

Current Status

The product is in early access as of 2026. The two-model critique (Deliberation) remains in beta while real-project dogfood and blinded evaluation continue. Medha's synthetic baselines are available; human-labelled release evaluation is still in progress. Early-access delivery is targeted within seven days of purchase. An iPhone and iPad version is available via TestFlight request. Korvo Agents — starting with Google Calendar, Drive, and Korvo Workspace integration — are announced as coming next, with a private cloud execution model and visible run history.