Skip to content

Supachad

An always-on agent that runs while you sleep. Sandboxed, budgeted, earns autonomy gradually. These docs cover the architecture, the runtime contracts, and the operational surface — what Chad is and how Chad works.

  • New here?

    Read the introduction for the 90-second tour. What Chad is, what makes it different from a chat agent, and the three properties that drive every architectural decision.

    Introduction

  • Architecture

    Three concentric isolation rings, four memory layers, two execution substrates. The full design with diagrams.

    Overview

  • Sub-agents

    Seven kinds, each pinned to a binary path under an L7 network policy. How spawning works and how to add a new kind.

    Orchestrator

  • Autonomy

    Every irreversible action runs through chad-action-gate. The autonomy boundary is a JSON file you can read.

    Autonomy

  • Memory

    Four stores, four shapes. Always-injected workspace files, semantic LTM, named-entity wiki, hybrid vector+graph brain.

    Memory stack

  • Operations

    Eleven standing crons. The schedule is the agent's contract with the day. Each cron is a budget-guarded one-line wrapper.

    Cron schedule

  • Premium escalation

    When Nemotron's depth runs out, Chad escalates to Anthropic Claude through a tightly gated wrapper. AuthContext, L7-pinned binary identity, weekly spend roll-up.

    Premium

  • Front-ends

    Open WebUI as the chat surface. Chad exposed as an OpenAI-compat model. Same gateway, same memory, same action gate as a cron.

    Front-ends

  • Runs IDE

    A web IDE for durable Smithers workflows at runs.supachad.com. Watch, edit, and launch runs with full history, crash-resume, and model fusion across the whole NVIDIA catalog.

    Runs IDE

  • Capabilities & boundaries

    What Chad does today, and what the design deliberately refuses to do. Current state, plus the reasoning behind each permanent "no".

    Capabilities & boundaries

Recent

The full sequence is in the changelog. The current state of chad-dev is what these docs describe — if a claim here disagrees with the source, the source wins; please file an issue.

  • Runs IDE at runs.supachad.com. A web IDE for durable Smithers workflows — run history, live tree, .jsx editor, model fusion. See Runs IDE.
  • Two orchestrators, bridged. chad-spawn (one-shot, GHA-isolated) plus chad-Smithers (durable workflows); runSpawn() offloads heavy workflow steps onto a GH runner. Five chad-spawn / cron features now run as crash-resumable workflows.
  • Nemotron 3 Ultra 550B + reasoning by default. Newest NVIDIA open model on the hosted profiles; the tool-call harness bug is absent in Ultra.
  • Async GHA spawns. chad-spawn --async, reconciler cron, branch retention, opencode kind, --binary-override flag.
  • NVIDIA fallback in the GHA runner. codex / opencode auto-route to integrate.api.nvidia.com when no OPENAI_API_KEY is set.

What this site is not

  • Not a NemoClaw user manual — that lives in the parent project's Sphinx docs. NemoClaw is the hardened sandbox blueprint Chad runs inside; these docs are Chad-specific.
  • Not a "run your own Chad" guide — Chad relies on private credentials (Proton, Anthropic, NVIDIA) and a private state repo. See Reproducing for the honest answer.
  • Not branded marketing. The voice is direct and technical because the audience is too.