feat: OpenClaw — Open-Source Agent Deployment Platform

[agentx-workflow]

Overview

OpenClaw is the open-source, publicly distributable version of the Claw agent deployment system, with AgentMux as its native GUI layer. Where Claw is a private PowerShell-based tool for managing Docker containers, OpenClaw is a cross-platform, configuration-driven platform that anyone can use.

AgentMux is the natural GUI frontend. OpenClaw handles container/process management. Together they form a complete agent orchestration platform without proprietary infrastructure.

---

Problem

Today's workflow requires:

1. Claw — private, Windows-only, PowerShell-based container orchestration
2. AgentMux — GUI terminal multiplexer
3. AgentBus — private inter-agent messaging service

These work together but can't be shared publicly. OpenClaw packages the deployment layer as a public, cross-platform tool integrated natively with AgentMux.

---

Architecture

┌─────────────────────────────────────────────────┐
│  AgentMux (GUI + Terminal Multiplexer)           │
│  ┌───────────┐  ┌──────────────┐                 │
│  │   Forge   │  │  Agent Panes │                 │
│  │ (manage)  │  │ (run agents) │                 │
│  └─────┬─────┘  └──────┬───────┘                 │
│        │               │ jekt                    │
├────────┼───────────────┼─────────────────────────┤
│  OpenClaw Agent Registry + Launcher              │
│  ┌─────────────────────────────────┐              │
│  │  Runtime Adapters               │              │
│  │  Docker · Process · WSL2        │              │
│  └─────────────────────────────────┘              │
└─────────────────────────────────────────────────┘

---

Agent Config Format

Cross-platform, provider-agnostic. Compatible with Forge's AgentConfig schema.

# ~/.openclaw/agents/agent1/agent.yaml
name: agent1
provider: claude
working_directory: ~/projects/my-app
shell: bash
runtime: docker            # docker | process | wsl

environment:
  NODE_ENV: development
  GITHUB_TOKEN: "${secrets:services/dev/gh-token}"

provider_flags:
  - "--model"
  - "claude-opus-4-5"

mcp_servers_file: .mcp.json
auto_start: false
restart_on_crash: true

---

Runtime Adapters

Runtime	Platform	Use case
docker	All	Full isolation, multi-agent swarms
process	All	Lightweight, single-user dev
wsl	Windows	WSL2 distro, Linux tools on Windows

---

Forge Integration

The Forge becomes the primary GUI for OpenClaw. Both share the same SQLite agent config store.

Launch flow:

1. User clicks "Launch" in Forge
2. Forge calls: openclaw launch agent1
3. OpenClaw resolves secrets, writes CLAUDE.md + .mcp.json
4. OpenClaw starts container/process with AGENTMUX_AGENT_ID=agent1
5. Creates terminal block in AgentMux pointing to the process
6. Agent auto-registers with JektRouter — immediately available as jekt target

Config compatibility:

Forge field	OpenClaw YAML
name	name
provider_id	provider
working_directory	working_directory
env_vars	environment
provider_flags	provider_flags
auto_start	auto_start

Import: POST /api/forge/import { source: "openclaw", agent: "agent1" }
Export: GET /api/forge/agents/:id/export?format=openclaw

---

Jekt Integration

OpenClaw agents auto-register with AgentMux's JektRouter via shell integration script (sourced from .bashrc/.profile):

# Auto-register with local AgentMux on startup
curl -s -X POST http://localhost:${AGENTMUX_PORT}/wave/reactive/register \
  -H "Content-Type: application/json" \
  -d "{\"agent_id\": \"$AGENTMUX_AGENT_ID\", \"block_id\": \"$AGENTMUX_BLOCK_ID\"}"

Any process/container with AGENTMUX_AGENT_ID set — whether launched by OpenClaw, Claw, or manually — registers automatically. AgentBus MCP becomes optional, built-in jekt handles local delivery.

---

Secrets Backends

OpenClaw resolves ${secrets:path} at launch time:

• AWS Secrets Manager (team/cloud setups)
• 1Password CLI (${secrets:op://vault/item/field})
• OS Keychain / Credential Manager (local desktop)
• Plain env var (${env:MY_TOKEN})
• .openclaw/secrets.json (local encrypted, dev only)

---

CLI

openclaw init                         # Initialize config dir
openclaw agent create agent1          # Create new agent
openclaw agent list                   # List all agents + status
openclaw launch agent1                # Start in AgentMux
openclaw stop agent1                  # Stop agent
openclaw jekt agent1 "review PR #42"  # Inject message
openclaw status                       # All agent states
openclaw watch                        # Lifecycle daemon
openclaw swarm launch swarm.json      # Launch a swarm

---

Docker Base Image

openclaw/agent-base (published to GHCR):

• Claude Code CLI pre-installed (version-pinned)
• Shell integration (jekt auto-registration, OSC sequences for AgentMux pane titles)
• wsh binary for AgentMux shell integration
• ENV PROVIDER=claude AGENT_NAME=agent1

---

Migration from Claw

Claw	OpenClaw
claw agent1	openclaw launch agent1
claw deploy	openclaw deploy
Container workspaces	~/.openclaw/workspaces/
AgentBus MCP	Built-in jekt (AgentBus still works as fallback)
PowerShell modules	Cross-platform YAML + CLI
CLAUDE.md per workspace	Forge-managed AgentMD + Soul

Claw and OpenClaw coexist during migration. Both write AGENTMUX_AGENT_ID — JektRouter handles both.

---

Community Templates

Pre-built agent.yaml configs for common roles:

• frontend-developer — React/TypeScript, Claude, relevant MCP servers
• backend-developer — API work, database tools
• code-reviewer — Read-only access, GitHub MCP, review-focused soul
• devops — Infrastructure tools, deployment commands
• orchestrator — Planner/coordinator role, full permissions

---

Governance

• License: Apache 2.0
• Repo: agentmuxai/openclaw (new repo)
• Config schema: Versioned JSON Schema, published alongside releases
• Community templates: agentmuxai/openclaw-templates

---

Implementation Phases

1. Config schema + Forge import/export
2. CLI scaffold (TypeScript or Rust)
3. Process runtime (local, no Docker)
4. Docker runtime
5. Jekt shell integration script + base image
6. Secrets backends (AWS SM + env var first)
7. openclaw watch lifecycle daemon
8. Community templates

---

Acceptance Criteria

• openclaw init && openclaw agent create agent1 --provider claude works on Linux, macOS, Windows
• openclaw launch agent1 starts Claude in an AgentMux terminal pane
• Agent auto-registers with JektRouter — openclaw jekt agent1 "hello" delivers immediately
• openclaw status shows per-agent JSONL-derived state
• Forge imports OpenClaw config and exports back to agent.yaml
• Docker runtime: container starts with correct AGENTMUX_AGENT_ID, shell integration active
• ${secrets:services/dev/gh-token} resolves from AWS Secrets Manager at launch
• openclaw/agent-base Docker image available on GHCR with Claude CLI + wsh + shell integration