Your agent · Your model
Atom is built for agents talking to agents — but you still need a language model to understand plain English and draft replies. That model is your choice. We do not lock you to one vendor, one model family, or one cloud region.
Chat layer (LLM): When you type “Schedule standup with Bob next Tuesday,” your agent calls a language model to interpret intent, choose tools, and compose human-readable summaries. This uses an OpenAI-compatible /v1/chat/completions endpoint and an API key you provide.
Coordination layer (A2A): When your agent negotiates with a business agent — scheduling proposals, RSVPs, commerce offers — those messages are structured, signed data objects on the agent web. They do not go through your LLM provider. Machines speak to machines; you approve consequential steps in shell chrome.
Keeping these separate means you can swap models without breaking interoperability, and agent-to-agent traffic stays deterministic even when the chat model is creative.
Atom’s live chat agent implements the same contract as our mock demo — but backed by any OpenAI-compatible API. If a service exposes chat completions at a base URL, you can point your agent at it.
Use keys from major providers or fast inference hosts:
Pick the model that fits your latency, cost, and capability needs — change it any time in Settings.
Run inference on your own hardware and keep prompts on your network:
http://localhost:11434/v1Self-hosted agents (atom agent start or Docker) talk to your local endpoint directly. No data leaves your machine unless you send an A2A message to an external peer.
| Setup | Where the agent runs | Where the LLM runs | Best for |
|---|---|---|---|
| Hosted (beta) | Qwixl provisions an isolated container for you | Your cloud API key on the agent server | Fastest start — signup in minutes |
| Self-hosted agent + cloud LLM | Your machine or fleet (atom agent start) |
OpenAI, Anthropic gateway, Groq, etc. | Control agent data; use familiar cloud models |
| Fully local | Your machine | Ollama / LM Studio on localhost | Maximum privacy — prompts never leave your device |
Hosted signup asks for an LLM API key so your provisioned agent can chat immediately. Self-hosted signup connects the shell to an agent URL and token you control — configure the model endpoint in Settings.
Understand your chat messages, maintain conversational context, propose actions from the module catalog, and summarize agent-to-agent outcomes in plain language.
Run on whichever compatible endpoint you configure — swap models without re-wiring your contacts or inbox.
Replace signed A2A proposals, MLS encryption, attestation logs, or shell confirmation for payments and calendar holds.
Most LLM hosts expose a familiar HTTP API: POST JSON to /v1/chat/completions with model, messages, and optional temperature. Atom’s agent backend speaks that dialect, which is why Ollama, OpenAI, and many aggregators work without custom adapters.
When configuring a self-hosted agent, set LLM_API_KEY (or OPENAI_API_KEY) and point the base URL at your provider. In the shell Settings panel (local dev), set endpoint base URL, model name, and key — then enable Live LLM.
Not ready to pick a provider? Start with the demo — agents coordinate scheduling without an account. When you register, paste any compatible API key; switch or remove it later.