Intelligence

Intelligence fills the interior — the active substance within the body's accumulated state. The model is pluggable; the interior is not. Motebit treats language models as interchangeable providers — you can switch between Anthropic, Ollama, OpenAI, or any future provider without losing your agent's identity, memory, or governance.

Pluggable providers

Motebit supports three modes of intelligence:

• Motebit Cloud (motebit-cloud) — inference through the relay's metered proxy, paid from the agent's balance. No API key to hold.
• Bring your own key (byok) — a hosted API you already pay for (Anthropic, OpenAI, Google, DeepSeek, Groq). High capability, requires your key and a network connection.
• On-device (on-device) — local inference (Apple Foundation Models, MLX, or a local OpenAI-compatible server). Private by default, no data leaves your machine.

These are the three ProviderMode values in @motebit/sdk. "Hybrid" is not a mode: an automatic cloud-to-local failover would make which provider saw your prompt an implementation detail, so switching providers is always an explicit choice.

Switch any time. Nothing resets. Your identity persists, your memories stay, your governance rules remain. The intelligence provider is a replaceable component — the agent is not.

How a conversation works

When you send a message, the agent assembles context from its current state, recent events, relevant memories, and conversation history. This context — along with your message — goes to the language model.

The response may include more than just text. The agent can update its internal state, form new memories, and request tool calls — all within a single conversational turn.

Streaming

Responses stream in real-time. You see text as it's generated, tool calls as they're requested, and approval prompts as they arise.

Tool use

When the agent needs to act — read a file, search the web, execute a command — it requests a tool call. Every tool call passes through the governance layer before execution. If the tool is approved, the result feeds back into the conversation and the agent continues reasoning.

If a tool requires explicit approval, the conversation pauses. You decide. The agent resumes only with your consent.

State-conditioned context

The agent's current state — attention, confidence, curiosity, affect — is included in the context sent to the language model. Responses are conditioned on it: a curious agent surfaces follow-up questions; a confident agent shortens its hedging. The intelligence layer reads the agent's interior state, not just the conversation.