OpenClaw
OpenClaw
One-line summary: Open-source always-on personal agent (formerly Clawdbot) whose Gateway heartbeat wakes on a timer, reads HEARTBEAT.md, and usually no-ops — the proactive counterpart to a queue drain.
Vintage: 2026-08. Community heartbeat docs and a 2026-08-15 @steipete post. Product names and cost tables move fast.
What it is
An open-source personal agent that sits on messaging channels and a local Gateway. The feature that matters for unattended-all-day-agent-loop is the heartbeat: every N minutes the agent checks pending work without a user prompt. Most cycles end HEARTBEAT_OK (nothing to do). Tasks live in a user-edited HEARTBEAT.md; every token in that file is loaded every cycle.
Formerly Clawdbot (the name used in Sam French's April 2026 writeup). Creator: Peter Steinberger (@steipete). This page tracks the loop architecture, not a product recommendation.
Why it matters to this thread
It is the documented always-on / cheap-check variant of the outer loop. Claude Code Routines and Cursor Cloud Agents start an isolated job. OpenClaw starts a check that should usually be free. A research bot that already has a durable queue wants the drain for work and (optionally) a cheap heartbeat for "is the queue empty?"
Key facts (from 2026-08-18-autoresearch-effective-all-day-bot-loops)
- Heartbeat interval default 30 minutes (
interval: 1800seconds) in the community docs. - Cost table (community docs, model prices as printed there): default Opus 4.6 + 30 min ≈ $10–20/day; Haiku 4.5 + 60 min + 8h quiet hours ≈ $0.60/day (claimed 96% reduction vs a ~$15/day default).
HEARTBEAT.mdshould stay short; a 500-token file × 48 cycles/day is 24,000 tokens of task definition alone.- Multi-agent setups can give each agent its own interval; community docs note a v2026.2.9 bug that fired all agents on the main schedule (fixed later).
- @steipete (2026-08-15): the team "moved … to build openclaw with openclaw"; "share agent sessions as URLs is a superpower."
Strengths
- Proactive without requiring a human to start a session.
- Cheap no-op path (
HEARTBEAT_OK) if the checklist is tight and the heartbeat model is small.
Weaknesses / concerns
- Heartbeat is "the single biggest cost driver for most OpenClaw users" if the model/interval are left at frontier defaults.
- Community docs, not a first-party Anthropic/OpenAI manual. Cost figures are estimates in that doc, not a measured bill.
- Always-on presence is a different job than draining a research queue with promote+ingest gates.
Open questions
- How much of the heartbeat pattern should a vault research bot copy versus leaving "is there work?" to
QUEUE.mdstatus?