v0.4 Aether 0.4 — multi-agent runtime is now stable →

Build, deploy & run
AI agents on your own GPUs.

Glixy Aether is an open-source AI assistant and multi-agent runtime — chat, plan, call tools, browse, and orchestrate sub-agents. MIT-licensed, MCP-ready, runs on Llama, Mistral, Qwen or any OpenAI-compatible endpoint.

⬇ Download for Windows ★ Try the cloud version ◐ View on GitHub

or one-line install: iwr https://glixylabs.com/install.ps1 | iex

★

GitHub stars

⑂

Forks

◐

Contributors

⬇

Weekly installs

⚙

MIT

License

aether · assistant llama3-70b · local

YOU

Find this week's open security tickets, summarize them, and draft a Slack update.

◈

⚙ tool · jira.search ok

✦ agent · summarizer ok

▣ tool · slack.draft running…

Found 7 open tickets. 3 are P1 (auth, vault rotation, audit log). Drafting a thread now

+ Ask Aether anything… ⌘ K

Multi-agent run · live 3 agents · 12 steps

◈

Orchestrator

planning · 4 steps

⚙

Researcher

8 sources · done

✦

Summarizer

running…

▣

Notifier

queued

Tokens14,820

GPU87%

Latency342 ms

What's inside

Everything you need to ship real agents

A batteries-included assistant + a programmable runtime to compose, schedule, and observe multi-agent workflows. All open source.

◈

Assistant UI, in your browser

Chat, plan, browse, and use tools — like a private ChatGPT, but pointed at your own models, your own data, your own keys.

Streaming Markdown Voice Files

$aether start --model llama3-70b → assistant ready · http://localhost:7777 → tools loaded · 14 mcp servers ✓memory synced · 2.1k notes

Launch the assistant →

⌬

Multi-agent runtime

Compose orchestrator → workers → critics with a few lines of Python. Plan, branch, retry, escalate — observable end-to-end.

Orchestrator Subagents Tracing Replay

◈

⚙ ✦ ▣ ⚡

Build a graph →

⚙

Tools & MCP

Call shell, web, code, files, Slack, Jira, GitHub — or any MCP server. Hot-reload tools without restarting the agent.

MCP Functions Shell

⛁

Memory that lasts

Short-term, episodic, and semantic memory layers. Agents remember tasks, preferences and prior runs — across sessions.

Vector SQLite Replay

📚

RAG, native

Drop in PDFs, codebases, sites, Notion. Hybrid search + re-rank, grounded answers with citations. No glue code.

Hybrid Re-rank Cite

🖥

Local LLMs, first

Llama, Mistral, Qwen, Phi via Ollama or vLLM. Or point Aether at OpenAI / Anthropic — your choice, same SDK.

Ollama vLLM OpenAI-compat

⏱

Workflows & schedules

Trigger agents on cron, webhooks, file changes or chat. Long-running jobs survive restarts and resume from the last step.

Cron Webhooks Resume

⛨

Privacy & control

Self-host on your laptop, server, or our GPUs. Encrypted memory, role-based tools, audit log. Your data never leaves you.

Encryption · AES-256

Self-host · 1 binary

Audit · Live

License · MIT

Python SDK

Compose agents with a tiny SDK

A few primitives — Agent, Tool, Memory, Run — get out of your way. Write Python, not YAML. Trace every step, replay any run, swap models with one line.

Single-file agents that scale to multi-agent graphs
Built-in tools: web, shell, files, code, browser, MCP
Streaming, structured output, schema-validated tool calls
Plug-in any vector DB — Weaviate, Qdrant, pgvector, Pinecone
OpenTelemetry traces · Prometheus metrics · replay debugger

★ Star on GitHub Read the docs

# multi-agent research crew with Aether
from aether import Agent, Tool, Crew

researcher = Agent(
  name="researcher",
  model="llama3-70b",
  tools=[Tool.web(), Tool.arxiv()],
)

writer = Agent(
  name="writer",
  model="qwen2-72b",
  memory="semantic",
)

crew = Crew(
  orchestrator="plan-then-delegate",
  agents=[researcher, writer],
)

result = crew.run(
  "Write a brief on India's AI ecosystem."
)
# → 12 steps · 14.8k tokens · 4.2s
# → trace: aether.dev/r/8f2a91

Architecture

How a single request becomes a multi-step run

From your prompt to a planner, to tools and sub-agents, to a grounded answer — every step traced.

Prompt · user

"Audit our infra for cost wins."

Planner · LLM

Tools + Subagents

web · shell · sql3 subagents

Answer · grounded

"Save ₹4.2L/mo by right-sizing 4 GPU pools."

Steps

Tools called

Tokens

14.8K

End-to-end

4.2 s

Use cases

Built for devs, researchers and teams

🧑‍💻

Dev copilots

Repo-aware coding agent with shell + git access

🔬

Research crews

Plan → search → read → synthesize, with citations

🎧

Internal support

Triage tickets, draft replies, escalate by SLA

📊

Data analysts

SQL agent with schema memory and chart tools

⚡

Ops & SRE

Cron-driven agents that watch logs and act

🛡

Compliance

Self-hosted, on-prem, your weights, your audit

Aether vs cloud agents

Open, ownable, and faster on your data

Capability

◈ Glixy Aether

Cloud agent platforms

Source code

✓ MIT, on GitHub

✗ Closed

Self-host on your GPU

✓ Single binary

✗ Vendor cloud only

Bring any model

✓ Llama, Mistral, OpenAI…

~ Limited list

MCP tools

✓ Native

~ Partial

Multi-agent graphs

✓ First-class

~ Add-on

Trace & replay

✓ Built in

~ Per-vendor

Per-token pricing

✓ $0 (your GPU)

✗ Marked-up tokens

Pricing

Free forever. Pay only for hosting.

Aether is MIT-licensed. We charge for managed hosting on Glixy GPUs and enterprise support.

Open Source

₹0forever

For builders, researchers and self-hosters.

★ Full source code (MIT)
★ Assistant + multi-agent runtime
★ All built-in tools + MCP
★ Local LLMs via Ollama / vLLM
★ Community support on Discord

View on GitHub →

Built in the open, with the community

Issues, PRs and discussions live on GitHub. New release every 2 weeks.

★

Stars on GitHub

⑂