ek_OpenClaw

Knowledge Share - OpenClaw build and integration.

OpenClaw VM Substrate

Minimal, Governed AI-Agent Runtime (v0.x)

Overview

This repository defines a minimal Linux VM substrate for running OpenClaw as a human-in-the-loop AI agent runtime. The design prioritises cost control, auditability, reversibility, and security over autonomy or scale.

This is not an autonomous agent platform. It is a controlled execution environment for experimenting with AI-assisted workflows under explicit governance.

[ Global_Substrate_Settings ] → [ Substrate_Compute ] → [ ek_OpenClaw ] (governance) (infra) (runtime)

---

Design Principles

• Human-initiated execution only
• No unattended or background agents
• Explicit cost boundaries (API-metered)
• Minimal persistent state
• Low attack surface
• Easy teardown and redeploy

“Under-automated” is a feature at this stage.

---

Toolset Summary (Current Build)

Compute / VM

• Platform: Cloud VM (Azure-backed in current build; cloud-agnostic pattern)
• Size: Small footprint (≈ 1–2 vCPU, ~8 GiB RAM, modest disk)
• GPU: None
• Role: Single-purpose agent substrate (not a desktop or general server)

---

Operating System

• OS: Linux (Ubuntu-class)
• Desktop: None (headless)
• Services: Minimal (cloud agent only)
• Networking: Private access (SSH); no public services exposed

---

Memory & Swap Configuration

• RAM validated post-provisioning (free -h)
• Disk validated (df -h)
• Swap enabled (~2 GiB) to:
  • avoid OOM kills
  • stabilise low-memory agent runs
• No aggressive tuning; defaults preferred for predictability

Goal: stability, not peak performance.

---

Terminal & Session Management

• TMUX installed
  • Persistent sessions across SSH disconnects
  • Clean separation of:
    • OS management
    • OpenClaw execution
    • diagnostics / logs

TMUX is treated as required tooling, not optional convenience.

---

Python Runtime

• Python 3 installed via OS packages
• Virtual environments (venv) used exclusively
• No system-wide Python packages beyond essentials
• Clean separation between:
  • OS Python
  • project/runtime Python

---

OpenClaw

• OpenClaw installed locally
• Executed in embedded (local) mode
• No long-running daemon
• No background agents
• No auto-execution

OpenClaw is invoked explicitly per session.

---

External API Connectivity

• External LLM APIs only (e.g. OpenAI, Anthropic)
• No Chat UI credential reuse
• Credentials supplied via:
  • environment variables
  • local auth profiles
• API usage is:
  • metered
  • revocable
  • bounded by provider spend limits

There are no implicit credentials and no hidden call paths.

---

Explicit Non-Goals (By Design)

This substrate intentionally excludes:

• Autonomous or self-directing agents
• Multi-agent swarms
• Cron jobs, triggers, or unattended execution
• Write access to production systems or IaC
• GUI desktops or public endpoints
• Heavy observability stacks (ELK, Prometheus, etc.)
• Unbounded “memory” or vector stores
• Broad or unrestricted tool execution

Each exclusion is an architectural decision, not an omission.

---

Repository Intent

This repository exists to support:

• Reproducible VM builds
• Deterministic agent execution
• Clear governance boundaries
• Incremental capability layering

Future additions must be:

1. Explicitly justified
2. Cost-bounded
3. Auditable
4. Easy to remove

---

Near-Term TODO

• Add & test a dedicated Chat channel
  • Interactive, short-lived sessions
  • No persistent memory by default
  • Basic logging and cost visibility
• Document:
  • session lifecycle
  • reset/teardown procedures
  • minimal redeploy steps

This will be the first packaged capability layered on top of the substrate.

---

Status

• Current state: Foundations complete (v0.x)
• Posture: Under-powered, under-automated, over-controlled
• Next milestone: First reproducible chat capability with cost visibility