Skip to content

Configuration

atomic-agent is built to be inspected and tuned. Configuration covers which model it talks to, how the prompt is budgeted, and which external providers and tools are connected.

Models & providers

By default the runtime uses a local llama-server. You can also configure:

  • External llama-server endpoints
  • OpenAI-compatible HTTP providers (text or embeddings)
  • OpenRouter-style providers

Cloud providers are used only when you deliberately configure them — the default path keeps everything local.

Prompt budget

The runtime keeps a stable prefix — persona, rules, tools, skills, capabilities, and instructions stay byte-stable within a session so the KV-cache (cache_prompt / slot_id) can be reused. The tail — conversation, memory, world state, recalled notes, lessons, procedures, and loaded skill bodies — is clipped into a predictable prompt budget.

Parallelism

Independent read-only calls can run concurrently after a single inference. The default is up to 4 parallel tool calls, with a hard ceiling of 16.

Approvals & traces

Dangerous operations pass through approval gates. Every prompt, decision, and result is written to append-only NDJSON traces locally, so you can replay prompt drift and audit exactly what happened.

Remote control

Pair a single Telegram account for remote control with inline approval buttons, or connect MCP servers to expose their tools, resources, and prompts through the same tool registry.