anything.network

architecture • February 28, 2026

The Technical Stack: How OpenClaw Actually Works

OpenClaw architecture diagram
OpenClaw's architecture is a reference implementation for what a modern agent gateway looks like: local-first, multi-channel, with typed capability execution.

Core Architecture

OpenClaw operates as a local-first gateway — a single control plane that manages sessions, channels, tools, and events. Built on Node.js 22, it runs on macOS, Linux, and Windows (via WSL2).

The platform has four integrated subsystems:

Gateway WS Control Plane — Manages sessions, presence, configuration, cron jobs, webhooks, and hosts both Control UI and Canvas.

Pi Agent Runtime — Operates in RPC mode with tool streaming and block streaming capabilities. This is where agent reasoning happens.

Session Model — Handles direct chats, group isolation, activation modes, queue modes, and reply-back. Each session is isolated with its own context window.

Media Pipeline — Processes images, audio, and video with transcription hooks, size caps, and automated temp file lifecycle management.

Model Routing

OpenClaw supports multi-provider model access with built-in failover:

Provider Model Input/1M tokens Output/1M tokens Best For
Anthropic Claude Opus 4.6 $5.00 $25.00 Highest quality
Anthropic Claude Sonnet 4.5 $3.00 $15.00 Production default
Anthropic Claude Haiku 4.5 $1.00 $5.00 High-volume, fast
OpenAI GPT-4o $2.50 $10.00 Alternative provider
Local Ollama + Llama 3.3 8B $0 $0 Zero-cost, private
The recommended configuration for most users is Claude Sonnet 4.5 — $15-25/month for moderate usage. Using Haiku cuts costs 60-70%. Opus doubles them.

Subscription economics matter: if monthly API costs exceed $20, a Claude Pro subscription is better value. At $100+/month, Max 5x delivers significant savings. At extreme usage, Max 20x at $200/month provides 5-10x cost reduction over equivalent API spending.

Multi-Channel Inbox

This is where OpenClaw's architecture becomes relevant to agent infrastructure design. The platform exposes a unified capability surface across:

Primary: WhatsApp, Telegram, Discord, Slack, Signal, Google Chat, Microsoft Teams, iMessage

Extended: Matrix, Zalo, WebChat

Services: GPT, Spotify, Philips Hue, Obsidian, Twitter, Browser automation, Gmail, GitHub, and 50+ integrations

Each channel is a typed wrapper in Anything.network terms — a versioned capability unit with defined input/output contracts. The agent runtime discovers available channels and composes routes across them.

Security-by-Default

OpenClaw implements DM pairing by default — unknown senders must complete a pairing process before the assistant processes their messages. Public inbound DMs require explicit opt-in. The openclaw doctor command surfaces risky configurations.

This is the correct security model for autonomous agents: deny by default, explicit allow. It maps directly to the governed network concept in Anything.network's capability model.

The $0 Stack

For zero-cost deployment:

Trade-off: slower responses (3-8 seconds vs <1 second with cloud APIs) and lower quality. But functional for personal assistants, testing, and budget-constrained deployments.

What To Do Next

Read Part 3: From SDR Tool to Capability Surface for how OpenClaw's patterns map to the Anything.network abstraction layer.