Atomic Agent

Atomic Agent (atomicagent.io) is an open-source, local-first AI agent that runs its control loop on your machine — driving your browser, files, shell, cron jobs, Telegram, and MCP tools. It ships with llama.cpp for local models and also accepts OpenAI-compatible cloud providers, so you can point it at Qubax and run any of 340+ models with full tool calling.

Provider config (~/.atomic-agent/config.json)

{
  "llm": {
    "activeTextProvider": "qubax",
    "providers": [
      {
        "id": "qubax",
        "kind": "openai-compatible",
        "baseUrl": "https://api.qubax.ai",
        "defaultChatModel": "claude-sonnet-5"
      }
    ]
  }
}

# Per-provider key file (chmod 600). Shell-exported vars win over .env.
QUBAX_API_KEY=qbx_live_YOUR_QUBAX_KEY

atomic-agent            # or: atag
# LLM tab → Providers → n (new provider) → OpenAI-compatible
#   Base URL: https://api.qubax.ai   (NO /v1 — Atomic appends it)
#   API key:  qbx_live_YOUR_QUBAX_KEY
# The wizard verifies the key with a one-token completion before saving.

atomic-agent models search "claude" --provider qubax --refresh
atomic-agent models search "vision tools" --json

# Switch models mid-session with /model inside the TUI

Setup steps

1. 1

   Install Atomic Agent

   Run curl -fsSL https://atomicagent.io/install | sh (macOS Apple Silicon, Linux x64/arm64; Windows has a PowerShell installer). Also available on GitHub at github.com/AtomicBot-ai/atomic-agent under the MIT license.

2. 2

   Open the provider wizard

   Start atomic-agent, open the LLM tab, choose Providers, and add a new OpenAI-compatible provider. Or edit ~/.atomic-agent/config.json directly with the block above.

3. 3

   Point it at Qubax

   Set baseUrl to https://api.qubax.ai — stored WITHOUT the /v1 suffix, because Atomic Agent appends /v1/... itself. Add your qbx_live_ key in the wizard (or in ~/.atomic-agent/.env as QUBAX_API_KEY).

4. 4

   Pick a model and run

   The picker pulls Qubax's live /v1/models catalog — pick any model (e.g. claude-sonnet-5) and start giving tasks. Use /model to switch models mid-session; risky actions stay approval-gated.