awlib

AnimaWorks — Organization-as-Code

Talk to your team like a messenger app. Behind the scenes, they ship software 24/7.

AnimaWorks is an open-source framework that brings together persistent AI agents (Digital Animas) — each with their own long-term memory, identity, role, and schedule — to operate 24/7 as a living organization.

On the surface, you interact with them through a familiar, LINE-style messenger chat UI. Behind the scenes, agents running in isolated OS processes decompose goals, write code in parallel git worktree sandboxes, run tests and multi-model reviews on each other's work, open Pull Requests, autonomously repair CI failures, and monitor post-deploy logs without human intervention.

A single message from you / GitHub webhook / Scheduled trigger (Heartbeat / Cron)
   ↓
Lead Anima decomposes & delegates (SQLite TaskStore, mutex keys, delegation contracts)
   ↓
Engineer Animas implement in isolated worktrees → Test → Multi-model PR review → CI self-repair → Deploy verification
   ↓
Only exceptions and approval gates escalate back to you via LINE-style chat, voice call, or notifications

AnimaWorks does not hardcode a rigid build pipeline. Instead, it provides organizational infrastructure: neuroscience-inspired long-term memory, process isolation, an atomic task board, permission & PII isolation (Enclave), and multi-channel communication. Because of that, the exact same organization can ship code, research competitors, handle customer inquiries, take meeting minutes, and coordinate over Slack or phone calls. It is not a workflow engine — it is an organization.

日本語 README | 简体中文 README | 한국어 README

---

Proven in Production

This is not a toy demo. For over six months, an 8-Anima organization powered by AnimaWorks has continuously developed, operated, and maintained a production SaaS product.

Metric (Mar–Aug 2026, aggregated from primary execution logs)	Production Result
Pull Requests authored autonomously by agents	302 PRs (267 merged)
Pull Requests operated by agents (review, CI repair, conflict resolution)	752 PRs (721 merged)
Share of tasks self-initiated & executed by the organization	99.7% (out of 31,215 total tasks, only 92 were human-initiated)
Tasks automatically generated from GitHub events (August alone)	2,508 tasks

The redesigned LINE-style chat UI: monitor team status from the left Anima rail, chat with green/white bubbles, reply with inline quotes, approve release notifications, download attached reports, and watch the live bust-up react in the bottom-right corner.

3D Workspace: real-time visualization of the org tree, each Anima's live status, and recent actions.

The Pixel Office is not a decorative animation — every speech bubble reflects a real task running right now in the background.

---

:rocket: Try It in 60 Seconds (Demo Environment)

No API key required if you already have Claude Code CLI installed or are logged into Codex.

Clone, set up with a one-liner, and launch a pre-populated 3-Anima organization (Product Manager kaito, Lead Engineer sora, Assistant hina in Japanese presets; alex, kai, nova in English presets) complete with 3 days of activity history. It runs in an isolated directory (~/.animaworks-demo) so it never touches your production ~/.animaworks data.

curl -sSL https://raw.githubusercontent.com/xuiltul/animaworks/main/scripts/setup.sh | bash
cd animaworks
uv run animaworks demo

Open http://localhost:18501 to start chatting with a team that is already at work (Demo Guide →).

4 Built-in Presets: en-business (English, realistic — default) / en-anime (English, anime) / ja-business (Japanese, realistic) / ja-anime (Japanese, anime). Pass --reset when switching presets (e.g., uv run animaworks demo --reset --preset ja-anime).

---

Quick Start (Build Your Own Organization)

macOS / Linux / WSL

curl -sSL https://raw.githubusercontent.com/xuiltul/animaworks/main/scripts/setup.sh | bash
cd animaworks
uv run animaworks start

Windows (PowerShell)

irm https://raw.githubusercontent.com/xuiltul/animaworks/main/scripts/setup.ps1 | iex
cd animaworks
uv run animaworks start

Windows Note: AnimaWorks uses Unix-specific APIs (fcntl, pty, os.killpg, etc.), so native Windows execution is not supported. Running setup.ps1 in PowerShell automatically installs WSL and sets up AnimaWorks inside your WSL Linux environment.

Open http://localhost:18500 in your browser and follow the setup wizard:

1. Language & Theme (en / ja / zh / ko)
2. LLM Connection (Claude Code subscription login, Codex login, API keys, or local Ollama)
3. Create Your First Anima (choose from templates such as dev-lead, full-stack, assistant, marketer, translator, or create a custom role)

You can also initialize and operate everything headlessly from the CLI without a browser:

uv run animaworks init
uv run animaworks start
uv run animaworks chat alice "Check our sprint goals for today and delegate tasks to the team"

Docker Compose / Manual pip Installation

Docker (Zero host pollution — recommended)

docker compose up launches an isolated Linux container with Node.js, GitHub CLI (gh), Playwright, Bubblewrap, and Claude Code / Codex / Gemini CLIs pre-installed.

git clone https://github.com/xuiltul/animaworks.git && cd animaworks
# Configure credentials in .env (ANTHROPIC_API_KEY / CLAUDE_CODE_OAUTH_TOKEN / GH_TOKEN, etc.)
docker compose up -d --build

Headless initialization without the browser wizard:

docker exec -it <container> animaworks init --skip-anima
docker exec -it <container> animaworks anima create --name alice --template dev-lead
docker exec -it <container> animaworks config set setup_complete true
docker exec -it <container> animaworks send <your-name> alice "Hello!"

Manual pip Installation (Python 3.12 / 3.13 recommended)

git clone https://github.com/xuiltul/animaworks.git && cd animaworks
python3 -m venv .venv && source .venv/bin/activate
pip install --upgrade pip && pip install -e ".[all-tools]"
animaworks start

---

On the Surface: A Team You Talk to Like a Messenger App

Many autonomous agent frameworks force humans to stare at raw log streams or wire up complex node graphs. AnimaWorks flips that model: day-to-day communication feels as effortless as LINE or Slack, while deep inspection of memories and background work is one click away.

1. LINE-Style Talk Screen, Left Anima Rail & Live Presence Panel

• Intuitive Bubble Timeline: Your messages appear in green bubbles on the right; Anima responses appear in white bubbles on the left, with smooth streaming previews of thoughts and tool executions.
• Left Anima Rail: Member avatars line the left edge, showing at a glance who is actively working (blue ring) and who is waiting for your reply (amber badge). Dismiss inactive chats with ×, re-open via +, split the screen into up to 4 panes, or launch a Meeting Mode (moderated group discussion with up to 5 Animas).
• Bottom-Right Presence Panel: A live bust-up portrait sits in the bottom-right corner of the chat, shifting expressions in real time (smile, thinking, troubled, surprised, etc.) and displaying a thought bubble of what the Anima is currently doing.

2. Human-Centered Feed & One-Tap Approvals / Reply Quotes

• Noise-Free Conversation Feed: In a 24/7 organization, hundreds of Heartbeat patrols, inter-Anima DMs, and Cron jobs happen behind the scenes. The chat UI automatically collapses background activity into compact single-line bundles so your conversation stays clean (toggle Background in the header anytime to expand everything).
• Approval Cards & Reply Quotes: When an Anima needs human sign-off (call_human), an interactive card with Approve / Reject / Comment buttons appears directly in chat. You can also quote-reply to any message or notification (Reply Chip) so the Anima knows the exact context of your instruction.

3. Bidirectional File Attachments, Inline Viewer & Slack Mirroring

• Drag & Drop Any File: Attach images, PDFs, Markdown, CSVs, source code, Office documents, audio, or video directly into chat.
• Outbound File Cards from Animas (attach_file): Animas attach generated reports, spreadsheets, and images directly to their replies. Click any attachment card to open the built-in Fullscreen File Viewer (sanitized Markdown preview, syntax-highlighted code, audio/video player).
• Slack Conversation Mirroring: Messages exchanged with an Anima on Slack while you are away are automatically mirrored into the Web chat timeline.

4. One-Tap Memory Popover & Fullscreen Voice Call

• D3 Entity-Connection Bubble Graph: Click an Anima's face icon or the Memory button in the chat header to open an anchored popover into their mind. Explore the interactive D3 entity-connections bubble chart (Connections), current work state (State), knowledge & procedures (Files), daily episodes (Journal), tool history (Activity), and scheduled routines (Routines).
• Fullscreen Voice Call (Call Screen): Click the phone icon to enter a fullscreen voice call powered by a low-latency front-lane model, VOICEVOX / Style-BERT-VITS2 / ElevenLabs / Irodori TTS, and streaming faster-whisper STT — complete with lip-synced, blinking bust-up animation and live captions. For critical production incidents, Twilio integration (core/phone/) even lets an Anima place an actual phone call to your mobile phone.

The Memory Popover opens with one click from the chat header or avatar, visualizing atomic facts, knowledge, procedures, journals, and an interactive D3 entity-connection graph.

Dashboard: oversee the live org chart, items needing attention, model quota consumption, process health, and real-time activity across the entire organization.

---

Under the Hood: A 24/7 Autonomous Organization Engine

Behind the friendly chat UI is a battle-tested multi-process runtime designed to run continuously for months without human hand-holding.

FastAPI Server (Thin Control Plane / WebSocket & SSE / Enclave Client)
  ├── MemoryService Process (Dedicated ChromaDB + Embeddings HTTP owner, SQLite WAL)
  ├── Anima Root Process: kaito (Manager)
  │     ├── Front Lane (Low-latency response / Voice / Routing)
  │     ├── Main Chat / Inbox / Heartbeat / Cron Lanes
  │     └── Task Runner Subprocesses (IPC v2 — isolated execution per heavy task)
  ├── Anima Root Process: sora (Lead Engineer)
  └── Anima Root Process: hina (Assistant)

1. Pick Up Right Where You Left Off — Multi-Layer Memory & Autonomous Maintenance

Unlike stateless chat wrappers that reset every session, each Anima has a filesystem-first, brain-inspired memory architecture: