Runtime

The call

Each run is one request to Claude (claude-opus-5 unless the manifest says otherwise) with:

1. the manifest's system prompt, cached as a stable prefix;
2. a second system block naming the builder and payout stock, so the model can answer "who made you";
3. the user's prompt as the single user message.

Adaptive thinking is on. Streaming is used so long answers do not time out. Server-side refusal fallbacks are enabled, so a policy decline re-runs on a fallback model inside the same call instead of returning nothing to a paying user. A refusal that survives the fallback is returned as text and is not charged.

The digest

digest = sha256(prompt + " " + output)

It is stored with the run and shown on the agent page. It exists so a run can be re-verified later without trusting neurodots's database.

Demo mode

Without ANTHROPIC_API_KEY, the runtime returns a labelled deterministic stand-in and still records the run and its reward, so the whole loop (run, accrue, settle) works on a fresh clone. Set the key in Vercel and the same button runs the real model.

Tools

tools is executable, not decorative. The local tools (chain, http, memory, clock) run on neurodots with your allow-lists enforced before the call, and they run even in demo mode: a chain read really reads Robinhood Chain. The server tools (web_search, web_fetch, code) execute on Anthropic's side inside the same call, so they need ANTHROPIC_API_KEY on the server; without it the trace says so instead of pretending.