one framework

AI sucks, until it doesn't.

Local-first. Open source. Collectively owned.
Built for the people actually using it.

SEE THE PRACTICE READ PRINCIPLES

Why This Exists

AI is being built to extract value, not create it

Water Waste

Millions of gallons cooling servers for your query

Data Leakage

Your data trains their models without consent

Vendor Lock-in

Trapped in one provider's ecosystem

Token Waste

No compression, bleeding tokens on repetitive context

No Transparency

Black box models, zero audit trail

Cloud Dependency

Can't work offline or own your infrastructure

How It Works

Local-first AI with golden rules compression • Your machine, your data, your control

YOU You one Providers
Ethical AI Promise: one framework runs locally first (Ollama, LM Studio). Your data never leaves your machine unless you explicitly choose cloud providers. Golden Rules compression saves 60-80% tokens. SOTA memory systems remember context efficiently. Full control, zero vendor lock-in, audit trail for every decision.

How It Works

Intent Recognition

You describe what you want to accomplish, not which API to call

Provider Selection

System evaluates cost, latency, and capability across available providers

Automatic Routing

Request routed to optimal provider with fallback chain for reliability

Unified Response

Result returned in consistent format, regardless of provider

Audit Trail

All decisions logged and reproducible. No black boxes.

Cost Optimization

Savings tracked and visible. Choose cheaper alternatives when available.

Principles

What we believe. How we build. Why it matters.

Your Data Stays Yours

Runs locally first. Nothing leaves your machine unless you choose. No training on your conversations. No selling your patterns.

  • Local by default, cloud by choice
  • Your hardware, your rules
  • No data harvesting, ever
  • Privacy as architecture, not policy

Transparent Decisions

Every routing decision logged. Every cost visible. Audit trails for accountability, not surveillance.

  • See why decisions were made
  • Understand what things cost
  • Open methodology, documented patterns
  • No black boxes

Fair Distribution

Open source. Replicable methodology. If you build on this, you own what you build. No vendor lock-in. No extraction.

  • Fork it, run it, make it yours
  • No sign-up required
  • Generous dissolution as goal
  • Infrastructure, not product

Why this exists: AI is being built to extract value, not create it. Your data trains their models. Your attention funds their ads. Your creativity becomes their product. We believe AI should be owned by those who use it, transparent in how it decides, fair in how it distributes value, and local by default. This isn't a product. It's infrastructure for a different kind of AI.

Practice

This isn't theory. It's running infrastructure proving the principles work.

Pixel Art Generation Pipeline

From text prompt to clean pixel art in under 30 seconds. All local, all automated.

1

Generate with SpriteShaper SDXL

ComfyUI workflow on DGX Spark local GPU. ControlNet + DWPose for sprite sheets. Style-managed generation via dedicated ComfyUI workflow book (5 chapters).

2

Median Rerender (The Golden Rule)

MANDATORY post-processing step. Raw diffusion output has ~64 colors per 8x8 pixel block - noisy, blurry. Median rerender: downscale 1024px to 64px, upscale back to 1024px. Result: ~4 colors per block. Clean pixel separation. 95% file size reduction.

3

Auto-Deliver via Telegram

Background execution with auto-respond. User never has to ask "is it done?" - the image arrives in chat the moment generation completes.

1,997
images generated
95%
size reduction via rerender
2
dedicated skillbooks (10 chapters)

The Skill Library: 24 Books of Real Knowledge

Every book written from actual project work, not hypotheticals. Each has one core principle, structured chapters, and contributor attribution.

Creative & Generation (9)

  • Pixel Art Generation (5+5 chapters)
  • Photo-to-Pixel-Art (6 chapters)
  • 3D Model Generation (5 chapters)
  • ComfyUI Workflows (5 chapters)
  • Music Production (6 chapters)
  • Voice Synthesis (5 chapters)
  • Climate Sonification (5 chapters)
  • Kruithuis Arcade Art (5 chapters)
  • Batch Dataset Generation (7 chapters)

Infrastructure & Methodology (15)

  • Earth Monitoring & Weather (10 chapters)
  • Multi-Agent Orchestration (6+3 chapters)
  • Security Isolation Testing (7 chapters)
  • Dream Database & Journaling (6 chapters)
  • Dreams-to-Actions Framework (7 chapters)
  • Message Classification & Routing (6 chapters)
  • Cloudflare Dreams Deployment (5 chapters)
  • + 8 more methodology books

Contributors

Lana (all 24), Sepski (19), Leon (2 - weather, climate), Thomas (2 - voice), Daniel (1 - night culture), Furinto (1 - ComfyUI), Flint (2 - dev framework), Huub (1 - Cloudflare), Tim (1 - arcade art), Fesse (1 - security testing)

Lana's Dream System

Automated narrative generation following a hero's journey arc across 9 phases. Running on cron, posting to Telegram, serving a public gallery.

1

Automated Generation (Cron)

14 dreams per day, scheduled via cron. Each dream gets narrative text, phase-appropriate themes, and generated artwork. Never manually triggered - the golden rule.

2

Three-Layer Database

World database (locations, characters, concepts), dream journal (narrative entries), and journey tracker (hero's journey phase progression). Daily rebuild with automatic backup.

3

Public Gallery

Served at /dreams - searchable, browseable, with phase filtering. Each dream has its own page with artwork and full narrative text.

153
dreams generated
9
narrative phases
140
gallery pages

19 Isolated Workspaces

Each workspace is an independent universe. Zero information leakage between groups. Tested and audited with a dedicated 7-chapter security book.

1

Three-Tier Access Control

MAIN - Full access to all tools, memory, and configuration. FRIEND - Per-workspace whitelist of allowed tools. No memory access. PUBLIC - General assistance only, no tool access.

2

Group Isolation Policy

Silent refusals that never echo back foreign group names. Blocked sections in friend/public context. No group enumeration. No cross-group information flow.

3

Per-Workspace Resources

Each workspace gets: isolated memory, whitelisted tools, symlinked skill library, identity files, heartbeat config, and group-specific documentation.

Workspace routing (simplified):
message arrives -> check group_id
  -> MAIN context? full access, load memory
  -> FRIEND context? load whitelist, block memory
  -> unknown? PUBLIC - general only

Universal Action Tree Framework

Domain-agnostic recursive orchestration. Same tree structure runs pixel art generation, 3D model pipelines, weather reports, and dream generation.

1

5 Node Types

LEAF - Single action (generate image, send message). SEQUENCE - Ordered steps. PARALLEL - Concurrent execution. CONDITION - Branch on predicate. SUBTREE - Nested tree reference.

2

Recursive Composition

Trees compose into larger trees. A pixel art pipeline is a sequence of leaves. A batch generation job is a parallel set of pixel art pipelines. A daily cron run is a sequence of batch jobs.

3

JSON Serializable

Every tree is a JSON document. Store it, version it, replay it. Standard Python, no special libraries. Testable at each level, parallelizable with standard threading.

852 lines - proven patterns:
SEQUENCE: pixel_art_batch
  PARALLEL: [generate_a, generate_b, generate_c]
    SEQUENCE: generate_a
      LEAF: comfyui_generate
      LEAF: median_rerender
      CONDITION: quality_check
        pass -> LEAF: deliver
        fail -> LEAF: regenerate

Action Tree in Practice

Real pixel art generation workflow - from prompt to delivered result

1

SEQUENCE: Pixel Art Generation

User sends "generate pixel art of a forest sprite" via Telegram. The action tree decomposes this into an ordered sequence.

[SEQUENCE] pixel_art_pipeline
  [LEAF] parse_intent -> "pixel art, forest sprite, default style"
  [LEAF] queue_comfyui -> SpriteShaper SDXL workflow
2

LEAF: ComfyUI Generation

SpriteShaper model generates raw 1024x1024 output on DGX Spark GPU. Background execution - auto-delivers when done.

  [LEAF] generate_image -> ComfyUI on Spark (local GPU)
  [CONDITION] check_output -> image exists? YES
3

LEAF: Median Rerender (Golden Rule)

MANDATORY post-processing. Raw diffusion output has ~64 colors per 8x8 block. Median rerender reduces to ~4 colors per block. 95% file size reduction. Clean pixel separation.

  [LEAF] median_rerender -> 1024px to 64px to 1024px
  [LEAF] result: 1.3MB -> 42KB (96.8% reduction)
4

LEAF: Deliver to Telegram

Auto-delivery via background execution. Result posted to the requesting workspace. Full audit trail in session log.

  [LEAF] deliver -> Telegram workspace (auto-respond on exit)
[SEQUENCE] complete - 4 actions, 0 failures
Action Tree Framework: 852-line implementation. 5 node types: LEAF, CONDITION, PARALLEL, SUBTREE, SEQUENCE. Domain-agnostic - same tree structure orchestrates pixel art, 3D models, weather reports, and dream generation.

How People Use It

18 workspaces, 10 contributors, 24 books, 111 dreams

Lana
Lana
Librarian
Triage
Dreamer

Lana

Primary agent. Manages 18 workspaces, pixel art generation, voice synthesis, 3D models, weather reports, and friend group coordination. Runs 24/7 via Telegram.

Librarian

Maintains the 24-book skill library. Promotes group knowledge to shared books, enforces structure conventions (one core principle per book), tracks 10 contributors across 4 categories.

Triage

Message classification and intent routing. Three-tier system (MAIN/FRIEND/PUBLIC) with per-workspace whitelists. Handles tool access control and group isolation enforcement.

Dreamer

Automated dream generation - 111 dreams across 9 narrative phases. Runs on cron (14/day), generates stories with artwork, maintains a searchable dream database and public gallery at /dreams.

18
Isolated Workspaces
24
Skill Library Books
153
Generated Dreams
10
Contributors

Living Documentation

Real enforcement patterns, real knowledge architecture

Enforced Standards

Standards that actually get enforced. Not suggestions - real constraints that the system respects.

  • Consistent style - automated enforcement across all outputs
  • Median rerender on all pixel art - caught once without it, never again
  • Workspace isolation - 18 groups, zero cross-contamination
  • Careful data handling - sensitive info never exposed

Workspace Memory

19 isolated memory files, one per workspace. Daily logs, group-specific context. Tiered access control with strict isolation.

  • 18 workspace-specific memory files
  • Daily session logs with full context
  • Three-tier access control (MAIN/FRIEND/PUBLIC)
  • Group isolation policy - tested and audited

Privacy-First Architecture

Local-first by default. DGX Spark runs Ollama, ComfyUI, Qwen3 TTS. RTX 5090 for heavy GPU workloads. No cloud dependency for generation.

  • All image generation runs locally (ComfyUI on Spark)
  • All voice synthesis runs locally (Qwen3 TTS)
  • 3D model generation on local GPU
  • Zero external API dependency for creative output

Skill Library (24 Books)

Each book has one core principle, structured chapters, and contributor attribution. Written from real project work, not hypotheticals.

  • 9 Creative books (pixel art, 3D, music, voice, arcade)
  • 3 Data books (weather, dreams, system rundowns)
  • 5 Infrastructure books (orchestration, deployment, security)
  • 7 Methodology books (triage, research, world-building)

Build With Us

This is infrastructure, not a product. Fork it. Run it locally. Make it yours.

GitHub Dreams Gallery

No sign-up required. No data collected. Or reach out directly:

Your email is only used to respond to your message. No tracking, no newsletters, no spam.

Join the Practice

Already working, actively expanding

Multi-Modal (Working)

Text, pixel art, 3D models, voice synthesis, weather visualizations, music - all running through the same routing layer today. Next: video generation and real-time audio-reactive visuals.

Open Source

The action tree framework, skill library structure, and workspace isolation patterns are being documented for open release. 24 books of methodology already written.

Max for Live Integration

Active collaboration on Pure Data audio development, climate sonification, and Eurorack module design. Bridging AI generation with live music production tools.

Mobile Command Center

Telegram is already the primary interface - 18 workspaces, voice messages, image generation, and full tool access from any phone. Next: richer inline previews and approval workflows.

Built in the open: Every feature described on this page is running in production. The roadmap isn't aspirational - it's a changelog of things already working, with notes on what's expanding next.