OpenClaw Agent Forge

Codex-built bootstrap for a live OpenClaw AI server and multi-agent operating environment.

This repository captures a reproducible local OpenClaw installation pattern, the operating notes behind it, and the guardrails for future AI agents.

What This Repository Tracks

• local launchers and helper scripts
• a bootstrap path for a fresh machine
• local memory structure and durable notes
• a Codex CLI-backed default model path without OPENAI_API_KEY
• tmux helpers for keeping the local gateway alive
• Telegram test-channel preparation notes
• lessons learned from the first installation pass

Published repository:

• https://github.com/SELLSYSTEMS/openclaw-agent-forge

It does not track the installed runtime or mutable state directories. It must also stay free of secrets because the repository is public.

Repository Layout

• AGENTS.md - instructions for future AI agents
• QUICKSTART.md - shortest path for a fresh operator or AI agent
• docs/install-runbook.md - step-by-step installation model
• docs/lessons-learned.md - mistakes and decisions worth preserving
• docs/shared-host-context.md - safe map of shared Codex, Node-RED, and host-level context
• docs/agent-topology.md - known agent folders, roles, and rediscovery rules
• docs/codex-cli-tui.md - how the shared OpenAI Codex CLI TUI works on this host
• docs/orchestrator-roadmap.md - recommended direction for orchestration, TUI, and cross-agent visibility
• docs/prompt-patterns.md - starter prompts for future agents on this host
• docs/telegram-test-plan.md - Telegram prerequisites and approval flow
• workspace/MEMORY.md - seeded public-safe memory for future installs and agents
• bin/openclaw-local - launcher with isolated OPENCLAW_HOME
• scripts/agent-landscape.sh - safe status snapshot of shared agents and services
• scripts/bootstrap-openclaw.sh - fresh setup bootstrap
• scripts/validate-local-setup.sh - smoke-test and validation
• scripts/start-gateway-tmux.sh - keep the gateway alive without systemd
• .github/workflows/smoke-check.yml - repo sanity checks on GitHub Actions
• memory/ - local Markdown memory vault

Local Layout After Bootstrap

• .openclaw/ - local OpenClaw installation prefix
• .openclaw-home/ - isolated runtime home
• workspace/ - default agent workspace
• memory/ - local Markdown memory vault

Bootstrap

/home/OpenClaw/scripts/bootstrap-openclaw.sh

Validate

/home/OpenClaw/scripts/validate-local-setup.sh

Run

/home/OpenClaw/bin/openclaw-local

Model Path

This setup is pinned to codex-cli/gpt-5.4.

• OpenClaw delegates agent turns to the installed codex CLI.
• Auth stays under the Codex CLI login state instead of this repo managing OPENAI_API_KEY.
• The gateway is configured for local mode on loopback and can be kept alive with tmux when systemd user services are unavailable.

Positioning

This repo is not the OpenClaw product source tree.

It is the operator repo around OpenClaw:

• bootstrap scripts
• isolated runtime layout
• memory conventions
• multi-agent workspace defaults
• installation lessons so the next AI agent does not repeat setup mistakes
• shared-host context so OpenClaw can coexist with other long-running agents and Node-RED automations

Shared Host Context

This machine has more than one active AI system.

• OpenClaw owns /home/OpenClaw
• the host also has a shared Codex CLI home under /root/.codex
• Node-RED runs locally under /root/.node-red and should be treated as shared host infrastructure available to all agents
• multiple terminal-driven Codex agents may already be active at the same time

The repo documents those relationships in a public-safe way. It should capture paths, rules, and operating patterns, but never passwords, tokens, or owner-specific identifiers.

Canonical Remote

git remote add origin https://github.com/SELLSYSTEMS/openclaw-agent-forge.git

Memory Strategy

The current best-practice default is file-based local memory in Markdown.

• It is transparent and local.
• It is easy for agents to inspect directly.
• It is easy to diff and back up.
• It avoids premature dependence on a vector service.

Obsidian is optional as a UI over the same folder. A vector database should be introduced only when the note corpus is large enough that keyword search is no longer effective.