The open-source coding agent you control

Many agents ship with hardcoded prompts and tools. Aether puts everything in config files you control, so every token in context is yours to mold.

Written in Rust, MIT licensed, batteries-included.

Docs GitHub

$ brew install contextbridge/tap/aether

$ curl -fsSL aether-agent.io/install | sh

$ cargo install aether-agent-cli

Run in a TUI, Editor/IDE or Headless

You’ll be up and running in two commands

Run in a TUI

First run auto-scaffolds config files, tools, and a default agent prompt.

$ aether
✓ Created .aether/settings.json
✓ Created .aether/mcp.json
✓ Created .aether/DEFAULT.md
✓ Created AGENTS.md
Launching TUI...

Run in your editor

Expose over ACP for VS Code, JetBrains, or any editor with agent support.

$ aether acp

Run headlessly

Run headless for scripts, CI, and automation.

$ aether headless "refactor auth"

Use any LLM you want, cloud or local

You can bring your own too.

Cloud

Anthropic, OpenAI, OpenRouter, DeepSeek, Moonshot, Zai, LlamaCpp and more are supported out of the box.

{
  "agents": [
    {
      "name": "Plan",
      "description": "Plans implementation strategy",
      "model": "codex:gpt-5.4",
      "userInvocable": true,
      "reasoningEffort": "high",
      "prompts": ["PLAN.md"]
    },
    {
      "name": "Build",
      "description": "Builds features and fixes bugs",
      "model": "zai:glm-5.1",
      "userInvocable": true,
      "prompts": ["BUILD.md"]
    },
    {
      "name": "Fast",
      "description": "Quick tasks and lookups",
      "model": "openrouter:minimax/minimax-m2.7",
      "userInvocable": true,
      "prompts": ["FAST.md"]
    }
  ]
}

Local

Run local models with llama.cpp or Ollama.

{
  "agents": [
    ...
    {
      "name": "Local",
      "description": "Runs locally via llama.cpp",
      "model": "llamacpp:unsloth/Qwen3.5-9B-GGUF",
      "userInvocable": true,
      "prompts": ["LOCAL.md"]
    }
  ]
}

Alloy multiple models together

Combine the strengths of multiple models by round-robining between them each conversation turn. You can even mix cloud and local models together.

{
  "agents": [
    ...
    {
      "name": "Alloy",
      "description": "Mixes cloud and local models",
      "model": "zai:glm-5.1,llamacpp:unsloth/Qwen3.5-9B-GGUF",
      "userInvocable": true,
      "prompts": ["ALLOY.md"]
    }
  ]
}

Control every token in your system prompt

Start with a blank slate and add only what you need.

Start with a blank slate

Agents begin with no system prompt or tools.

$ aether show-prompt -a Build

---
Prompt chars:            0
Tool schema chars:       0
Est. tokens:     ~       0
MCP tools:               0

Use any markdown files you want

Your system prompt is just an array of markdown files. Load AGENTS.md, CLAUDE.md, GEMINI.md...whatever and compose all of them together. @-file inclusion and shell command interpolation are supported out of the box

{
  "agents": [
    {
      "name": "Build",
      "description": "Builds features and fixes bugs",
      "model": "zai:glm-5.1",
      "userInvocable": true,
      "reasoningEffort": "high",
      "prompts": ["BUILD.md", "AGENTS.md"]
    }
  ]
}

# Build Agent
You are a staff+ level engineer.
...

# Agents
Use the Explorer agent to search code.
...

View your agent's system prompt with a single command

Including an est. token count

$ aether show-prompt -a Build

# System Prompt
You are an expert senior Rust engineer.
...

# Agents
Use the Explorer agent to search code.
...

---
Prompt chars:          120
Tool schema chars:       0
Est. tokens:     ~      30
MCP tools:               0

Connect tools via MCP servers (local or remote)

Be minimal…or maximal. Give your agent a single tool (Bash) and point it at some CLIs + skill files. Or, go full batteries-included by adding a few lines to your .aether/mcp.json file.

Batteries included MCPs

Aether includes several built-in MCP servers you can optionally enable. They run in a custom in-memory transport.

coding — file I/O, grep, bash, and web fetch
lsp — language-aware symbol lookup, diagnostics, and rename
skills — reusable domain knowledge as slash commands
subagents — spawn parallel child agents with their own models and tools
tasks — persistent task tracking across turns
survey — allow the agent to ask you questions with structured output
plan — submit markdown plans for approval and feedback

{
  "agents": [
    {
      "name": "Build",
      "description": "Builds features and fixes bugs",
      "model": "zai:glm-5.1",
      "userInvocable": true,
      "prompts": ["BUILD.md", "AGENTS.md"],
      "mcps": [".aether/mcp.json"]
    }
  ]
}

{
  "servers": {
    "coding":    { "type": "in-memory" },
    "lsp":       { "type": "in-memory" },
    "skills":    { "type": "in-memory" },
    "subagents": { "type": "in-memory" },
    "tasks":     { "type": "in-memory" },
    "survey":    { "type": "in-memory" },
    "plan":      { "type": "in-memory" }
  }
}

Any MCP server

Plug in any external MCP server via stdio, SSE, or streamable HTTP. Remote server OAuth credentials are stored securely in your keychain.

{
  "servers": {
    ...
    "linear": {
      "type": "http",
      "url": "https://mcp.linear.app/mcp"
    }
  }
}

MCP tool proxy stops tools from eating your context

Aether can wrap external MCP servers in a built-in tool proxy, so agents discover nested tools on demand instead of loading every schema up front.

{
  "servers": {
    "proxy": {
      "type": "in-memory",
      "input": {
        "servers": {
          "chrome-devtools": {
            "type": "stdio",
            "command": "npx",
            "args": ["-y", "chrome-devtools-mcp@latest"]
          },
          "linear": {
            "type": "http",
            "url": "https://mcp.linear.app/mcp"
          }
        }
      }
    }
  }
}

Ready to build?

Use the full CLI or pick individual crates — aether-core, llm, mcp-servers, wisp, crucible — to compose your own agent.

Read the docs View on GitHub