A self-evolving, personalized OpenClaw

The more you use it, the better it understands you.

• [2026/03/24] 🔥 Fully automatic setup — The plugin auto-starts the SCOPE server, auto-detects OpenClaw's model configuration, and auto-installs Python dependencies on first launch. No .env, no manual setup needed!
• [2026/03/24] 🔥 Added custom SCOPE prompts & domains tailored for personal AI assistants — user preference learning, code quality analysis, and communication style optimization!
• [2026/03/23] 🚀 EvolveClaw v1 released! Self-evolving prompt system for OpenClaw with zero code modification — plugin + sidecar architecture powered by SCOPE.

EvolveClaw turns OpenClaw into a self-improving agent that continuously adapts to your workflows. Powered by SCOPE (Self-evolving Context Optimization via Prompt Evolution), it observes how you interact with the agent — your tasks, your tool usage patterns, your preferences — and automatically evolves the system prompt with personalized guidelines that make the agent increasingly effective for you, not just any user.

No two EvolveClaw instances are the same. Over time, each one develops a unique personality shaped by its owner's habits.

Key adaptation: SCOPE was originally designed for task-specific benchmarks (e.g., HLE). EvolveClaw extends it with custom prompt templates and domain categories tailored for a personal AI assistant — focusing on user preference learning, task quality, communication style, and workflow patterns instead of domain-specific problem-solving heuristics.

SCOPE also support agent workflows like EvoFabric

• 💡 Why EvolveClaw?
• ⚙️ How It Works
• 🚀 Quick Start
• 🧬 What Makes It Self-Evolving
• 🏗️ Architecture
• 🎯 Design Decisions
• 🔌 API Endpoints
• 📖 Citation

Today's AI agents ship with a static system prompt — every user gets the same instructions. But users are different: some prefer terse answers, others want detailed explanations; some rely heavily on search tools, others write code directly; some use the agent for coding, others for research or writing.

EvolveClaw closes this gap with three core ideas:

1. Observe — The plugin captures your full interaction trace: model output, tool calls, tool results, errors, and the semantic nature of your task
2. Learn — SCOPE analyzes each trace and synthesizes a guideline if warranted — e.g., "When this user asks for refactoring, prefer small atomic commits over large rewrites"
3. Classify — Each guideline is classified as tactical (task-specific, ephemeral) or strategic (cross-task, persisted to disk as part of your personal memory)
4. Inject — On the next turn, all active guidelines are injected into the system prompt, structured by priority (strategic > tactical)
5. Forget — When you start a new session, tactical guidelines are cleared. The agent remembers who you are (strategic), not what you were doing yesterday (tactical)

This creates a virtuous cycle: the more you use the agent, the better it understands your preferences, and the more personalized its behavior becomes.

EvolveClaw requires a running OpenClaw instance. If you don't have one yet:

# Requires Node ≥ 22
npm install -g openclaw@latest

# Run the onboarding wizard (sets up gateway, workspace, channels)
openclaw onboard --install-daemon

For more details, see OpenClaw Getting Started. You'll also need Python ≥ 3.10 for the SCOPE sidecar server.

Option A — Install from registry (recommended):

# From npm
openclaw plugins install evolveclaw

# Or from ClawHub
openclaw plugins install clawhub:evolveclaw

openclaw gateway restart

Option B — From local source (Linux / macOS / WSL2):

git clone https://github.com/JarvisPei/EvolveClaw.git
cd EvolveClaw
./scripts/install-plugin.sh
openclaw gateway restart

git clone https://github.com/JarvisPei/EvolveClaw.git
cd EvolveClaw

# Add plugin path to OpenClaw config
$config = "$env:USERPROFILE\.openclaw\openclaw.json"
$cfg = Get-Content $config | ConvertFrom-Json
if (-not $cfg.plugins) { $cfg | Add-Member -NotePropertyName plugins -NotePropertyValue @{} }
if (-not $cfg.plugins.load) { $cfg.plugins | Add-Member -NotePropertyName load -NotePropertyValue @{} }
if (-not $cfg.plugins.load.paths) { $cfg.plugins.load | Add-Member -NotePropertyName paths -NotePropertyValue @() }
$cfg.plugins.load.paths += (Resolve-Path .).Path
if (-not $cfg.plugins.entries) { $cfg.plugins | Add-Member -NotePropertyName entries -NotePropertyValue @{} }
$cfg.plugins.entries | Add-Member -NotePropertyName evolveclaw -NotePropertyValue @{enabled=$true} -Force
$cfg | ConvertTo-Json -Depth 10 | Set-Content $config

openclaw gateway restart

The plugin handles everything automatically on first gateway startup: it auto-installs Python dependencies (scope-optimizer, fastapi, etc.) if missing, then auto-starts the SCOPE server. No manual pip install or server launch needed.

After restarting the gateway, check the logs for these lines:

[plugins] evolveclaw: activated (server=http://127.0.0.1:5757, agent=openclaw-agent, inject=append_system, maxGuidelines=30)
[gateway] evolveclaw: installing Python dependencies...        # first launch only
[gateway] evolveclaw: Python dependencies installed            # first launch only
[gateway] evolveclaw: SCOPE server started successfully
[gateway] evolveclaw: auto-configured SCOPE with OpenClaw's <your-model>
[gateway] evolveclaw: loaded N strategic rule(s), guidelines.length: 1

If you see activated and SCOPE server started successfully, EvolveClaw is running. Start chatting — the agent will begin learning from your interactions after every few exchanges.

Tip: If you prefer to manage the server process yourself (e.g., for custom .env settings or running on a different host), start it manually before OpenClaw and set "autoStartServer": false in the plugin config. The plugin detects a running server via /health and skips spawning.

{
  "plugins": {
    "entries": {
      "evolveclaw": {
        "enabled": true,
        "config": {
          "serverUrl": "http://127.0.0.1:5757",
          "agentName": "openclaw-agent",
          "injectMode": "append_system",
          "maxGuidelines": 30,
        }
      }
    }
  }
}

Unlike static prompt engineering or manual rule files, EvolveClaw implements a self-improvement loop with the following components:

evolveclaw/
├── package.json               # npm/ClawHub package manifest
├── openclaw.plugin.json       # Plugin manifest with config schema
├── plugin/                    # OpenClaw TypeScript plugin
│   └── src/
│       ├── index.ts           # Plugin entry: lifecycle hooks, guideline management
│       ├── scope-client.ts    # HTTP client for SCOPE sidecar
│       └── types.ts           # Shared type definitions (config, API, guideline metadata)
├── server/                    # SCOPE sidecar HTTP server (Python)
│   ├── server.py              # FastAPI server: step analysis, tactical reset, stats
│   ├── config.py              # Server configuration (env vars)
│   ├── prompts.py             # Custom SCOPE prompts & domains for personal assistant use
│   ├── requirements.txt
│   └── .env.template          # Environment variable template
└── scripts/
    ├── start-server.sh        # Start the SCOPE sidecar
    └── install-plugin.sh      # Auto-configure OpenClaw to load the plugin

• Feedback loop — No auto-feedback from the plugin (OpenClaw has no user_feedback hook). Could be added once SCOPE supports guideline removal or OpenClaw adds a feedback hook.
• Migrate to focused SDK subpaths — OpenClaw v2026.3.22 introduced openclaw/plugin-sdk/* subpaths (e.g., openclaw/plugin-sdk/plugin-entry) as the recommended import surface. Our plugin uses the monolithic openclaw/plugin-sdk which is still fully supported. Once installed via ClawHub, the focused subpaths can be adopted.

• SCOPE — The prompt evolution framework powering EvolveClaw
• OpenClaw — The AI agent platform

@software{pei2026evolveclaw,
  title={EvolveClaw: Evolving OpenClaw's System Prompt via Self-Improving Guidelines},
  author={Pei, Zehua and Zhen, Hui-Ling},
  url={https://github.com/JarvisPei/EvolveClaw},
  year={2026}
}

@article{pei2025scope,
  title={SCOPE: Prompt Evolution for Enhancing Agent Effectiveness},
  author={Pei, Zehua and Zhen, Hui-Ling and Kai, Shixiong and Pan, Sinno Jialin and Wang, Yunhe and Yuan, Mingxuan and Yu, Bei},
  journal={arXiv preprint arXiv:2512.15374},
  year={2025}
}

MIT