Aether

自托管多 Agent 工作台（本地版）

A fully local, self-hosted multi-agent workbench. Fixed workflow: Plan → Research → Execute → Review, with HITL approval, artifact storage, and configurable models.

Quick Start

npm install
npm run dev

Open [REDACTED]

What you get (no limits, fully local)

• Projects linked to local git repos (or current dir)
• Todos with layered execution: Plan → Confirm → Build → Review → Merge
• Multiple agents (Chief, Planner, Builder, Reviewer, custom)
• Multiple providers: Ollama (local), OpenAI, Anthropic, Google
• Git worktrees for safe isolated execution
• Conversation + steerable steps
• All features unlocked forever

First run

On first load it auto-creates:

• A "Personal" team
• A "Local Machine"
• Four starter agents

Configure models

Go to /app/agents

For pure local:

1. Install Ollama
2. ollama pull llama3 (or codellama, qwen2.5-coder, etc.)
3. Add agent with provider "ollama"

For cloud:

• Paste API keys when creating agents (stored only in your local data/store.json)

Architecture notes

This implementation realizes the Aether vision using a local-first approach:

• Step kinds: plan, implement, review, chief + revisions
• Worktrees at .workspaces/conv-*
• Branch naming tds/conv-* (or configurable)
• Agent assignment per step
• Tools: shell, repo ops, push, web fetch

Data lives in data/store.json. Worktrees in .workspaces/.

No cloud, no accounts, no machine limits, no paid tiers.

Run

npm run dev

License

Private / unspecified.