CowAgent vs ZeroClaw

CowAgent

MIT software with seven releases in 90 days, memory kept as plain files in the owner's workspace and three per-session permission modes. Source and Docker installs default to full-access, no action asks for approval, and an open report of 31 July 2026 says group members could reach the owner's files through the bot.

ZeroClaw

Supervised autonomy is the default, so tools outside a short read-only list wait for the owner's approval, and secrets are encrypted in the config file. The docs state that command execution is not written to an audit log, and the project publishes no privacy policy or terms.

CowAgent

1. Set agent_permission_mode to workspace-write or read-only before connecting any channel. Source and Docker installs default to full-access
2. Keep the bot out of group chats that include other people. Any member who mentions it reaches the same tools as the owner
3. Set web_password before changing web_host to 0.0.0.0. The password is empty by default
4. Set external_api_token to enable POST /v1/chat/completions, and send conversation_id to keep one session. The model field does not select a model
5. Run it in a container for a hard boundary. The docs and the source both say the permission modes are not isolation

ZeroClaw

1. Keep [risk_profiles.<alias>] level = "supervised". full runs every tool not listed in always_ask, and the YOLO preset removes the sandbox and path rules too
2. Add sending and account tools such as send_via, google_workspace and http_request to always_ask so a session-level Always grant cannot skip the prompt
3. List senders in [peer_groups.<name>] external_peers. ["*"] accepts anyone on that channel
4. Set [web_fetch].allowed_domains and [http_request].allowed_domains. Fetches run unprompted and the OS sandbox does not block egress
5. Run one-shot tasks with zeroclaw agent -a <alias> -m "<task>". A cron or webhook turn has no prompt, so shell calls fall to the command policy