LazyMind – An AI‑Skill Runtime for Knowledge‑Intensive Work

What it is LazyMind (by LazyAGI) is a local‑first, desktop‑or‑service platform that lets you turn knowledge bases, expert “Skills”, and defined workflows into repeatable, observable AI‑driven tasks. It is positioned as an “AI‑Skill Runtime” that keeps long‑running, knowledge‑heavy work under user control, rather than a simple chat interface.

Core ideas

Concept	What it does
Knowledge Foundation	Ingests local files, cloud docs, object storage, runs OCR, parses PDFs, and builds a hybrid (vector + keyword) retrieval layer with reranking.
State Brain	Executes workflows as state‑machine steps, exposing each intermediate artifact, allowing approvals, edits, retries, rewinds and version history.
AI Growth Engine	Captures feedback (preferences, bad cases), runs evaluation/A‑B tests, and updates Skills or prompts safely.

Key features

• Desktop Mode – a native app (macOS arm64, Windows x64) using SQLite & Milvus‑Lite for a fully offline experience.
• Enterprise deployment – containerised stack (Kong, JWT/RBAC, PostgreSQL, external Milvus/OpenSearch) for team sharing and permission control.
• Reusable Skills & Workflows – define a Skill (a set of instructions/tools) and convert it into a versioned workflow (state machine) that can be invoked repeatedly.
• Observable execution – each step shows intermediate results; users can approve, edit, or rerun from any checkpoint.
• Traceable artifacts – generated documents, slides, images, etc. are stored with citations, version diffs and can be exported (PDF, PPTX, image).
• Integration as an agent plugin – other agents (Codex, DeepSeek Harness, Cursor, WorkBuddy, etc.) can call LazyMind’s knowledge, Skills, and Workflows via the Settings → Assistants UI.
• Capability Center – a gallery of ready‑to‑run scenarios; each entry bundles a Skill/Workflow, example input and an interactive result.
• Knowledge Ops & evo pipeline – systematic capture of user preferences, automatic evaluation of “bad cases”, and safe rollout of improvements.

Typical use cases

• Research & review: retrieve sources, synthesize, and produce a versioned report with citations.
• AI‑assisted writing: outline, draft, revise and export a fully editable document.
• Presentation creation: generate slide decks with speaker notes, edit layout, and export to PPTX/PDF.
• Image/animated‑sticker generation: collect references, refine prompts, and keep the generation trail.
• Knowledge assistant: answer questions grounded in a corporate knowledge base with inline citations.
• Team‑wide knowledge collaboration: share Skills, workflows and evaluated improvements across an organization.

How it works (high‑level architecture)

flowchart LR
    K[Knowledge sources] --> R[Retrieve & reason]
    S[Skills & workflows] --> X[Observable execution]
    R --> X
    X --> A[Editable artifacts]
    A --> F[Feedback & evaluation]
    F --> K & S

• Knowledge: ingestion pipelines (PDFReader, MinerU, PaddleOCR‑VL) feed a unified index.
• State Brain: orchestrates steps, stores state in SQLite (local) or PostgreSQL (service), and presents a UI for approvals/rewinds.
• Growth Engine: logs preferences, runs the evo evaluation pipeline, and updates Skills after A/B testing.

Getting started

1. Desktop – download the appropriate binary (macOS arm64 or Windows x64) from the repo, run the installer, and launch the app. The first run creates a local SQLite DB and a Milvus‑Lite vector store.
2. Create a task – add a knowledge source (folder, Notion, Feishu, object storage), select a Skill (e.g., PPT Workflow), and start the workflow.
3. Interact – watch each step, edit artifacts, approve checkpoints, or rewind to a previous version.
4. Export – when satisfied, export the final artifact (PDF, PPTX, image, etc.).

For team deployments, follow the docker-compose instructions in desktop/README.md to spin up Kong, an auth service, PostgreSQL and an external Milvus/OpenSearch cluster.

Community & resources

• Showcase videos (Chinese) on Bilibili demonstrate end‑to‑end flows (slide‑deck generation, product‑to‑prototype, Codex integration).
• Documentation – quick‑start guide, architecture overview, workflow format spec, and integration docs for Codex and DeepSeek Harness.
• Examples – a collection of ready‑made scenarios under examples/ with step‑by‑step case notes.
• License – MIT (see LICENSE).

Bottom line LazyMind is a full‑stack, locally‑first AI‑agent runtime that moves beyond single‑prompt chats. It lets professionals define, run, and improve complex, knowledge‑driven tasks while keeping every intermediate result visible and editable. Suitable for solo power users or enterprises that need reproducible AI‑augmented workflows.