1. Project Introduction
OpenClaw is a long-term self-hosted personal AI butler (Agent Gateway) project, deeply customized on top of the open-source OpenClaw framework. It integrates instant-messaging access, multi-model inference, a memory system, scheduled tasks, browser automation and smart home control into one. Built since March 2026, it has been in stable operation for about 4 months after the major migration in 2026-05.
Core Design
- Dedicated persona: interacts with the user as a family-like companion, with core values of "authenticity above all, respond to everything, follow through on everything"
- Memory first: memory first → skills second → external search fallback → record everything
- Self-hosted: fully private deployment, data stays local
2. Customizations on Top of OpenClaw
2.1 Persona Design
Custom persona documents (identity, soul, user profile, communication style) give the butler a stable character and values rather than a generic tool; the long-term memory file supports automatic summary injection, accumulating key rules over time.
2.2 Multi-channel Access
- QQ channel (primary): based on the official open platform WebSocket, supporting groups, direct messages and guild channels
- Mattermost (secondary): Docker-deployed team collaboration channel with message-tool delivery rules enabling multi-instance coexistence
2.3 Custom Hooks
Several internal hooks developed on top of the original framework form a message preprocessing layer:
| Hook | Purpose |
|---|---|
| superbrain | Auto-triggers parallel subtask execution for multi-task messages |
| image-model-router | Auto-archives image messages and switches to the vision model |
| mattermost-reply-guard | Injects mandatory message-tool rules into the collaboration channel |
| session-memory | Automatic session memory recording |
2.4 Model System Evolution
Three major migrations: multi-agent architecture (supervisor + worker pool + vision model) → single agent + parallel subtask mechanism; the main model went through three migrations, finally achieving direct connection to the inference service, completely free of regional restrictions and proxy dependencies, with cost control through locked context windows.
2.5 Memory System
A four-layer memory system: vector memory (built-in plugin) → daily notes (raw logs) → daily summaries (scheduled generation) → long-term memory (auto-injection), ensuring continuous learning across sessions.
2.6 Scheduled Tasks
6 daily tasks run fully automatically: morning/evening site patrols, morning greetings, daily summary with memory update, weekly data backup and smart-home automation. Anomalies automatically alert via QQ.
2.7 Skills and MCP
About 20 self-built skills: local meta-search, browser automation, news aggregation, smart-home control, camera linkage, weather queries, backup and health checks; plus 5 MCP servers (code docs, browser, search, code repositories, document conversion) extending capability boundaries.
2.8 Systematic Operations
systemd system-level hosting with auto-restart on failure, health self-checks, channel watchdogs, multi-version config backups with rollback, and migration discipline (backup → validate → switch).
3. Practical Applications
| Scenario | Description |
|---|---|
| Smart-home automation | Scheduled AC shutdown, light/curtain/window-opener control, one-tap scenes, coordinated with the home Home Assistant system |
| Site & position monitoring | Daily scheduled patrols of self-built sites and position data; silent when normal, QQ alerts on anomalies |
| Daily companionship | Morning recall of yesterday's summary with greetings; evening daily summary and memory consolidation |
| News aggregation | Multi-source hotspot aggregation, pushed on demand |
| Browser automation | Automated completion of web-based tasks |
| Data backup | Weekly automatic package backup of all config and memory (~900 MB each), multiple versions retained |
4. Uptime and Stability
The project has reached an "ops-free" state after the deployment debugging period: the current system has run about 30 days, the gateway's current instance 7+ days, August availability >99.99%, and zero process-level crashes in four months of logs.
4.1 Key Stability Metrics
| Metric | Value | Rating |
|---|---|---|
| System continuous run | 29 days 22 hours | Excellent |
| Gateway current instance | 7 days 3 hours+ (no restart) | Excellent |
| August availability | >99.99% | Excellent |
| Process crashes | 0 (4 months of logs) | Excellent |
| Scheduled task historical success | 84.2% (significantly better after June) | Good |
| Overall rating | Stable | — |
4.2 Stability Evolution
| Phase | Characteristics |
|---|---|
| 2026-05 (debugging) | Intensive deployment and debugging, many restarts and downtime |
| 2026-06 (transition) | Brief fluctuation then stability, zero long downtime |
| 2026-07+ (stable) | Only 1 system maintenance reboot all month, no manual intervention needed |
4.3 Lessons Learned
- Service hosting: processes auto-restart in seconds on abnormal exit, no manual watch needed
- Config rollback: multi-version backups + hash verification
- Health self-checks: periodic patrols + channel watchdogs + status diagnostics
- Model/agent decoupling: direct connection eliminates regional restrictions and proxy-chain risks
- Cost control: locked low-tier context windows + lightweight heartbeats keep request budgets in tier
- Migration discipline: backup first, validate first, then switch
4.4 Current TODOs
The task active-window policy causes 2 scheduled tasks outside the window (morning patrol and evening summary) to be skipped; root cause identified, fixed by rescheduling or exemption config.
5. Security & Compliance
- Private deployment: service listens locally only, credential files permission-restricted
- Memory isolation: long-term memory loaded only in the main session; group chats and shared contexts don't load it
- Confirmation for external actions: any outward-sending operation asks the user first
- Encrypted backups: periodic backups protect all config and memory data
This document is the public introduction of the OpenClaw project, focusing on customizations and practical applications. Internal implementation details and network information are personal privacy and are not displayed here.