Skip to content

Introduction

Chad is an always-on assistant. Not a chat agent — those exist when you're typing at them and don't exist when you aren't. Chad exists in between. Chad reads mail every hour, drafts replies, watches its own budget, dreams its knowledge brain at night, and proposes self-improvements once a week.

Three properties define the design:

1. Always on

Chad is a process tree, not a tab. The agent's contract with the day is eleven standing cron jobs, each a one-line wrapper invocation that exits in under a minute. Heavy work happens inside the wrapper (Python, shell, detached nohup jobs); the model only sees "do the thing the wrapper does" because re-tokenizing a long prompt on every cron fire is expensive on a free-tier inference endpoint. Every cron runs with --light-context enabled, which strips skill and plugin frontmatter from the bootstrap and typically cuts per-fire input tokens by 70–80%.

The cron schedule is published here. It is deliberately conservative — fewer, cheaper runs beat reactive chat-bot behavior at this layer.

2. Delegates safely

Anything beyond a one-shot reply is a sub-agent spawn. Seven kinds today: coder, researcher, writer, reviewer, fitness, codex, opencode. Each kind is a YAML manifest that pins a binary path and a network policy preset.

Pinning the binary at the L7 proxy is the load-bearing safety mechanism. A reviewer kind is allowed to GET on api.github.com but not POST — meaning a compromised reviewer cannot post a malicious review even if its prompt is hijacked. A writer kind cannot reach GitHub at all. The principle is sub-agents draft; parents publish.

Spawns can route to one of two execution substrates:

  • local — sub-agent runs in Chad's container under the kind's L7 policy. Synchronous. Default for kinds that need access to Chad's local brain, embedded inference, or source clone.
  • gha — sub-agent runs on a GitHub Actions runner per spawn. Async-capable. The runner is fresh, unrelated to the parent's filesystem. Default for kinds whose work is self-contained (codex, opencode).

3. Earns autonomy

Every irreversible action — sending mail, posting a comment, modifying a cron, deleting a memory — passes through chad-action-gate. The gate reads auto-actions.json, which maps {action_type, target} pairs to one of three modes:

  • auto — Chad performs the action immediately
  • draft — Chad parks the action for human review
  • block — Chad must not perform the action

Anything not listed in the policy falls through to a default of block. New senders, new repos, new external services all start blocked and earn promotion to auto only after the operator demonstrates trust.

The kill switch is a single file: touch /sandbox/.openclaw/workspace/.auto-disabled halts every autonomous action regardless of policy. rm it to resume.

How Chad differs from a chat agent

Dimension Typical chat agent Chad
Lifecycle Per-session Always running
Trigger Human types something Cron, mail, GitHub issue, webhook
Memory Conversation history Workspace files + lancedb LTM + wiki + gbrain
Trust Implicit (you initiated it) Explicit policy file with per-target rules
Scope One task per turn Discrete spawns with budgets
Failure Retry the prompt Drafts the action; waits for review

What Chad is not

  • Not a chatbot. There's a chat surface (OpenWebUI), but the primary work happens unattended.
  • Not multi-tenant. One operator account, one mailbox, one GitHub identity. Multi-Chad scheduling is a future direction.
  • Not reproducible without infrastructure. Chad needs Proton, Anthropic (optional), NVIDIA inference, and a private state repo. See Reproducing for the full picture.
  • Not unsupervised. The autonomy roadmap is gradual on purpose. New action types and new senders go through human review before they auto-anything.

Where to go next

  • Architecture overview → architecture.md — the three rings, the spawn flow, the substrate split.
  • Memory stack → memory.md — what each store is for and where each fact belongs.
  • Orchestrator + sub-agents → orchestrator.md — kinds, manifests, prompt templates, how to add a new kind.
  • Substrates → substrates.md — local vs GHA, provider routing, async mode.
  • Autonomy → autonomy.md — the action gate, the policy file, the kill switch.