OpenClaw Quickstart 🦞

A drop-in library that turns a fresh OpenClaw install into a reliable, self-driving agent in three commands — an objectives-driven autonomous loop, a status dashboard, self-improvement, and safe cron management. Built to run on a modest base model (GPT-5 class), so it favours determinism and grounding over cleverness.

1. Install OpenClaw and add at least one model (normal openclaw onboard).
2. Install this bundle:

   openclaw skills install git:tunc-d/openclaw-quickstart@main
   

3. Tell OpenClaw: "run the quickstart setup".

It interviews you up front (models, budget, cadence, permissions), then sets everything up unattended and verifies it. Go AFK; come back to a working agent.

• Objectives loop. You add objectives; each must carry a machine-checkable Definition of Done (shell commands that exit 0). A deterministic gate cron (no LLM cost when idle) dispatches one bounded LLM step at a time toward the active objective, proves progress with the done-checks, and stops. When the queue empties, the loop disables itself.
• Plan-approval per objective. new-objective drafts a plan + done-checks; approve-objective queues it. The loop then runs unattended until done/blocked.
• Anti-garbage guards. Evidence-before-done, one step per tick, hard timeout, stall detector (auto-block after N no-progress ticks), budget cap, concurrency lock, append-only audit log, hard pause kill switch.
• Self-improvement. A LESSONS.md the worker reads every tick; the retrospective step appends concrete lessons after failures.
• Model policy. MODELS.md says which model for which task; the agent reads it and applies --model per task. It does not rewrite openclaw.json.
• Dashboard. A static page (regenerated from state) at http://127.0.0.1:18789/__openclaw__/canvas/quickstart/dashboard.html showing objectives / in-flight / blocked / done / crons / budget / log. Talk to the agent in built-in WebChat.
• Cleanup. Loop self-disables when idle; one-shot crons auto-delete; isolated sessions auto-prune (sessionRetention: 24h); stale locks are reclaimed.

SKILL.md          entry: "quickstart-setup" (run this)
setup/00..06      ordered, idempotent, verify-after-each setup checklist
templates/        seed state files (objectives, state.json, progress, LESSONS, MODELS)
skills/           the 6 operational skills installed into ~/.openclaw/skills
scripts/          gate.mjs · render-dashboard.mjs · lib-state.mjs (deterministic, Node stdlib)

• Why a command-cron gate instead of an LLM loop? Idle ticks must be free and safe. Plain code decides whether to spend any model at all; the model only runs for one bounded step when there is approved, in-budget, non-stalled work.
• CLI flags vary by OpenClaw version. Setup confirms each flag via openclaw <cmd> --help before relying on it and stops-and-asks on mismatch rather than guessing.
• No built-in objectives dashboard exists in OpenClaw — this ships a minimal static one rather than a bespoke web app, for reliability on a small model.
• Local-only by default — no outbound notifications; check the dashboard.