Rust Native · 3-Model Terminal Agent · v0.3.0

Plan. Build. Route.
One AI agent in your shell.

Aether runs three cooperating models inside a high-speed terminal loop: a planner maps the steps, a builder executes against code, and a router adapts to reality. Single static binary — sub-25MB RAM, zero telemetry, zero runtime dependencies.

GitHub
● 269 tests passing · ● 19 LLM providers · ● <25MB idle RAM · ● Free Zen provider
● aether v0.3.0 · zen
Aether live terminal demo showing 3-model planning, building, and routing in TUI

THE AGENTS

Three models. One loop.

Most agents run a single model. Aether runs three that hand off to each other — and the router rewrites the plan from what actually happened.

01

Plan

Reads the request, maps the repo, and writes a concrete step list — files to touch, commands to run, order of operations.

# plan
plan "add a --json flag to sync status"
▸ 3 steps
  1. parse flag in cli.rs
  2. emit json in sync.rs
  3. test round-trip
02

Build

Executes the plan against the codebase: edits files, runs commands, observes the results, and reports back what changed.

# build
build step 1/3
▸ edit crates/aether-sync/src/lib.rs
  + --json flag wired to Status
  ✓ cargo check passes
03

Route

Compares the result to the plan. Done? Hand off to you. Drifted? It rewrites the plan and sends it back to build — no dead ends.

# route
route result vs plan
▸ 1 deviation found
  json output missing tests
▸ plan v2 → build

The router is the difference: it treats the plan as a hypothesis, not a contract. Execution always feeds back into planning.

FEATURES

Everything a coding agent should be.

Built in Rust, tuned for low RAM, honest about providers.

01

Zero setup

Run aether and start chatting. Ships with the free zen provider — OpenAI-compatible, no API key, no account, no credit card.

02

19 providers

zen, openai, anthropic, google, deepseek, openrouter, ollama, groq, mistral, xai, cerebras, togetherai, fireworks, perplexity, moonshot, minimax, huggingface, lmstudio, github — plus any OpenAI-compatible endpoint.

03

Light on RAM

A native Rust binary idles in tens of MB — not hundreds. No Electron, no bundled Chromium, no interpreter overhead.

04

Sessions

Every chat auto-titles and saves. Browse, resume, or delete with aether sessions. Plain JSONL — grep it, script it, own it.

05

Recall

Cross-session memory: search every past conversation with aether recall "phrase" and jump straight back in.

06

MCP server

Aether speaks the Model Context Protocol over stdio and Streamable HTTP. Your editor and your agents share one context layer.

07

Sync across devices

Sessions sync to a private GitHub gist or any synced folder — merged line-by-line with origin fingerprints, so concurrent edits never silently overwrite.

08

Resilient transport

Streaming requests retry with exponential backoff and full jitter on 408/429/5xx — and the system prompt layers per-project overrides via .aether/system-prompt.md.

INSTALL

One command to first answer.

curl -fsSL https://aetherdz.github.io/aether-site/install.sh | sh

POWER USER

COMMANDS

The whole CLI.

FAQ

Questions, answered.

Three models. One loop.

Install it. Point it at a repo. Watch the router re-plan.

Star on GitHub

built in Rust