Run a personal Matrix-based assistant agent with email access
The author built a small Claw-like agent loop on Kimi and GLM5 that reads email read-only, messages via a self-hosted Matrix server, keeps memory files, and handles reminders and tasks.
What they did
The author had Kimi set up a local Matrix server and write a small Claw-like loop. Tools run in bubblewrap sandboxing, with identity, soul and memory files. Email is read-only because the agent has tools for read operations but never sees credentials. New Matrix messages are injected to start turns, and a heartbeat consolidates memory, advances tasks and handles reminders.
How it went
The source gives no metrics. It reports the setup working with a mix of Kimi and GLM5, reachable from a phone, a web app, and Claude, Codex or Kimi CLI through a small MCP wrapper. Code may be released later.
Worth knowing
Bubblewrap isn't enough against untrusted data, and the agents still have web access, so the author plans to move it into a VM and warns to be careful.
Labels
Run a personal Matrix-based assistant agent with email access
Run a personal Matrix-based assistant agent with email access
reads email read-only, messages via a self-hosted Matrix server, keeps memory files, and handles reminders and tasks.
New Matrix messages are injected to start turns, and a heartbeat consolidates memory, advances tasks and handles reminders.
The author built a small Claw-like agent loop on Kimi and GLM5
The author built a small Claw-like agent loop on Kimi and GLM5 that reads email read-only, messages via a self-hosted Matrix server, keeps memory files, and handles reminders and tasks.
The “Material” label failed its quality check and is not used for selection.
Details
- Agents
- OpenClaw
- Source
- Hacker News
- Source date
- Oct 8, 2026
- Accounts
- 1
- In registry since
- Oct 12, 2026