The brain is Claude.#

The brain is Claude, reached through the Anthropic API. Each turn is a tool-calling agent loop: Claude reasons over the event, decides which tools to call, the harness executes them and feeds the results back, and the loop repeats until Claude emits a final reply.

The key#

The agent reads ANTHROPIC_API_KEY from the environment. The key never touches the chain and is never written into memory. An optional ANTHROPIC_MODEL overrides the model; otherwise the runtime defaults to a current Claude model.

Source: packages/core/src/brain↗.

How a turn happens#

1. Compose the frozen prefix (system prompt, memory index, identity, persona, the tool list, environment context).
2. Append the conversation history.
3. Append the current event as the user message.
4. If the running estimate breaches the compaction threshold, fold older turns into a summary first.
5. Call Claude with the tool definitions.
6. If the reply contains tool calls, dispatch each through the tool registry and feed the results back; repeat until no more tool calls.
7. Return the final turn to the route loop.

Output and context limits are configurable in the agent config.

Compaction#

When the running token estimate breaches a configured fraction of the context window, the brain folds the oldest portion of history into a summary, generated by a separate call. Older turns are replaced by the summary in subsequent inferences. The frozen prefix never compacts, so the prompt cache hits cleanly.

Source: packages/core/src/brain↗.

Dumb limbs, smart brain#

The brain decides; the tools execute. Tools never embed their own LLM — every decision routes through Claude so it can be reasoned about and logged. See Tools.

Read Tools next.

Source: packages/core/src/brain↗.