Manifesto
Disclosure & Manipulation
Most AI tools show you the answer and hide everything else. Aether does the opposite.
It's a local-first, multi-provider agentic studio (Anthropic, OpenAI, Gemini, Ollama and compatible endpoints), built for one thing: to understand how an LLM agent really works — and to learn by using it.
Disclosure. An agent is usually a black box: you see the output, not the why. Aether opens the box — the assembled system prompt of every request, the context, the reasoning steps, every tool call with its output, the tokens. You no longer guess what drives the answers: you observe it.
Manipulation. You understand by doing. Breakpoints on tools, gates with a preview before irreversible actions, context injection, control from the CLI. Change a variable and see the effect: that's how you learn.
Aether is for people who work with AI and want to stop guessing at prompts — to really understand System Prompt, Tools, Skills and Gates, and the best practices that emerge from building agentic software.
What you understand with Aether
Agents, taken apart
System Prompt
What the model receives
The instructions, context and tools the model actually receives — usually invisible. In Aether you see the assembled system prompt of every dispatch: you understand how framing changes the answers.
Tools · MCP
What it can do
The functions the agent can call: filesystem, shell, any MCP server. You connect them in 1 click and see every tool call with its output, distinct from the rest: you understand when and why it uses a tool.
Skills & agents
How it's composed
Reusable skills and procedures. You compose skills, sub-agents and swarms — there's even skill-smith, an agent that helps you write new ones. You learn to structure capabilities instead of repeating prompts.
Gates & approvals
Who stays in control
Control over irreversible actions. You set tools to auto or gate and approve with a diff preview: you learn to let the agent run without losing control of it.
Lessons from the field
Best practices, not theory
- Brainstorm before code. An agreed, test-driven plan beats intuition. Using Obra's Superpowers skills is no longer optional: Aether's own agents would want to start from this rule, and all of Aether is built this way.
- Isolated sub-agents. Narrow task, narrow context: parallel work is reliable when you don't share too much context — anything extra is noise in the session.
- Gate before touching the world. File, shell and git are embedded tools by default (integrated via MCP) and pass through an approval with a preview: the LLM is free up to that point.
- Context is a resource. Per-workspace rooting, conversation forks and targeted tools — instead of giving everything to everyone everywhere.
What's still missing
The roadmap
- AST — read code as a syntax tree, not as text: reason about program structure, not strings.
- LSP — hook up Language Servers: diagnostics, definitions and references from the IDE, fed straight to the agent.
- Ingestion & RAG — index documents and codebases and retrieve only the relevant pieces as context, instead of pasting everything.
- Additional tools — more embedded tools beyond filesystem, shell and git.
- Multi-provider integration tests — end-to-end tests across multiple LLM providers, to guarantee behavioral parity.
- Remote control — drive the daemon and sessions remotely, not just from localhost.
- LDAP authentication — login via a corporate directory, for team use.
Appetite comes with eating.
In action
Aether live
Real screenshots: an Ollama and Anthropic agent working under your gaze.
A closer look
Details
skill-smith — a sub-agent that writes skills with you. You see (and edit) its system instruction: sub-agent → system prompt → skill.
