AnimaWorks runs a team of persistent AI agents (Animas) as an organization with a reporting hierarchy. Given a goal, a manager Anima breaks it into tasks and delegates them, engineers implement and test in isolated git worktrees, reviewers check each other's pull requests, and the agents fix CI failures, resolve conflicts, and deploy. They escalate to a human only when stuck or when a decision exceeds their permissions.
The framework itself does not hardcode that software pipeline. It turns GitHub events (CI failures, review comments, @bot commands, merge conflicts) into tasks, serializes work per pull request through exclusive keys, and orchestrates multi-model review, where each configured model reviews a PR and an integration task consolidates the verdicts. The agents do the rest with git, tests, and role-specific working conventions, so the same organization also handles email replies, meeting minutes, and Slack posts.
Each Anima runs as its own OS process with a schedule and file-based long-term memory. A heartbeat loop observes, plans, and reflects, cron jobs handle reports and monitoring, and nightly consolidation turns episodic memory into knowledge. Memory recall combines vector search, BM25, atomic facts, and an entity registry, and a forgetting pass marks low-activity memories for archival review rather than deleting them.
Different Animas can run on different engines: Claude Agent SDK, Codex CLI, Cursor Agent CLI, Gemini CLI, Grok Build CLI, or LiteLLM for providers such as OpenAI, Google, Mistral, AWS Bedrock, xAI, DeepSeek, and local models through Ollama or vLLM. The project compares itself with CrewAI, LangGraph, OpenClaw, and the OpenAI Agents SDK, positioning itself as an organization of autonomous agents rather than a role-based crew, a graph workflow, or a personal assistant.
A web dashboard shows a live organization chart, chat with any Anima (including a meeting mode for up to five), Slack-style channels, a task board, and an activity timeline, plus 3D and pixel-art office views driven by the live event stream.
Features
- GitHub event gateway: converts CI failures, review comments, bot commands, and merge conflicts into deduplicated agent tasks
- Task delegation: managers assign tasks with acceptance criteria, and tasks sharing an exclusive key run sequentially while others run in parallel
- Multi-model review: one review pass per configured model, consolidated into a single approve or request-changes verdict
- Autonomous schedule: heartbeat, cron jobs, and background tasks run each Anima around the clock, optionally on a cheaper background model
- Long-term memory: automatic recall across six parallel channels, hybrid search, consolidation, and reviewed forgetting
- Role templates: engineer, manager, writer, researcher, ops, and general roles with their own prompts, permissions, and default models
- Seven engines: Claude Agent SDK, Codex CLI, Cursor Agent CLI, Gemini CLI, Grok Build CLI, LiteLLM, and an assisted mode, with per-engine fallback
- Layered security: trust-boundary labeling of external data, command blocklists and allowlists, per-agent file sandboxes, and HMAC-checked webhooks
- Messaging: receives from Slack, Chatwork, Discord, and Zoom, and notifies humans over Slack, Chatwork, Discord, LINE, Telegram, or ntfy
- Voice chat: browser push-to-talk or hands-free voice with faster-whisper speech recognition and several speech engines, including ElevenLabs
- Avatar generation: builds full-body art, expressions, and 3D models from each Anima's personality through NovelAI, fal.ai, Meshy, or local Diffusers
