The PDF that showed up on its own
The school calendar arrived as a blurry image inside a Gmail message. The parent forwarded it straight to their OpenClaw agent and asked it to pull the dates, holidays, and early release days. By evening the events sat in the family calendar with invites already sent.
I had set this up after getting tired of typing the same schedule every year. OpenClaw read the image, extracted the details, and used the calendar tools without any further prompting. The whole thing stayed alive because the agent could wait for the next email and keep its state across days.
Claude Code would have finished the first parse and then gone silent. There is no way for it to keep listening or hold the calendar data without starting a fresh session.
How the state actually stuck around
OpenClaw writes everything to a plain-text MEMORY.md file that the agent reads and updates on its own. A 30-minute heartbeat also runs in the background and posts status to Discord or Telegram when you configure it.
I set the heartbeat to check pending tasks and log the current state every half hour. That loop let the agent watch for new Gmail messages or monitor device status without me sending a single prompt.
The same approach handled RAG over a thousand-page car restoration manual and watched the OctoPrint feed from the 3D printer. None of that would have survived past the first session with Claude Code.
The idle cost that showed up in the logs
I checked usage on a day with no new messages. The heartbeats alone were burning through tokens at the default interval. It added up to roughly $35 per inactive day even though nothing had changed.
One setup I tested ran stably on a Raspberry Pi Model B with free-tier Gemini. It handled Gmail, Google Drive, and GitHub tasks for 15 straight days. The persistence worked, but the background polling still showed up on the bill when I switched to a paid model.
Multi-agent setups with Anthropic models pushed some users past $700 a month once heartbeats and duplicated polling sessions started stacking.
When the updates started breaking the harness
Releases after 4.23 often pushed the node process to 100% CPU or hung the gateway. I kept an older instance ready because the new versions introduced orphaned sessions and constant restarts.
Rolling back took one command:
npm install -g openclaw@2026.4.23
The heartbeat settings lived in openclaw.json and looked like this:
"heartbeat": {"every": "30m", "target": "discord", "prompt": "# Heartbeat Checks..."}
These breaks forced me to use Codex or Claude Code just to repair the harness itself. The frequent changes created more maintenance than I expected.
The home setup that kept running
Once the version was locked, I expanded the agent into a full home butler. It controlled HVAC, shades, lights, music, and TVs through device integrations. A RAG system over the car manuals ran in the workshop with natural language STT/TTS.
File monitoring on the 3D printer via OctoPrint sent alerts to Telegram when prints finished or failed. OpenClaw's ClawHub skills and multi-channel connections made the whole loop possible. Claude Code has none of those external service ties.
The video clipping workflow
A content creator built a skill that processes one video each day into clips with hooks. It delivers the options via WhatsApp for selection, auto-uploads the chosen clips to TikTok, Instagram, and YouTube Shorts, then pulls weekly performance metrics to improve the next set of hooks.
The skill runs through ClawHub without daily input. OpenClaw handled the daily orchestration and external platform calls. For writing the actual video processing logic, Claude Code or Codex would have been the better starting point.
Where the coding agents win
OpenClaw's default PI harness uses fewer tokens than Claude Code's orchestration for simple persistent tasks. It falls apart on complex multi-step problem solving. That is why many people embed Codex as a sub-harness when code work appears.
Community reports are consistent. OpenClaw excels at Gmail parsing, device control, and document RAG. Claude Code and Codex stay superior for any code generation or editing.
The maintenance overhead that came with it
OpenClaw's open-source updates create ongoing troubleshooting that managed coding agents avoid. I spent time fixing gateway hangs and version conflicts instead of building new automations.
Embedding Codex for the coding parts still meant managing two separate contexts and their own state. The hybrid approach worked but never felt as seamless as a single specialized tool.
Here is how the options line up in practice:
| Criteria | OpenClaw | Claude Code |
|---|---|---|
| --- | --- | --- |
| Task focus | Persistent automation and integrations | Deep code editing |
| --- | --- | --- |
| Memory model | Persistent files + heartbeats | Session context |
| --- | --- | --- |
| Hosting | Self-hosted with flat compute | Managed API sessions |
What I settled on
For anything that needs to keep running across days without manual restarts, OpenClaw's harness is the practical choice. The MEMORY.md file and scheduled checks let it handle email triggers, device loops, and document parsing on its own.
For code generation, editing, or when you want to skip version headaches and idle token burns, Claude Code or Codex works better on its own. I still switch between them depending on whether the job needs persistence or sharp coding in the moment.