How Clawdbot (OpenClaw) Powers a Local‑First AI Agent Runtime with Dynamic Orchestration

Clawdbot (renamed OpenClaw) is a local‑first AI agent runtime that tightly integrates large language model (LLM) capabilities with the host system, toolchain, and communication platforms.

Core Architecture

Gateway (daemon) : Handles external channels (WhatsApp, Telegram, HTTP), maintains WebSocket connections, and manages authentication. It serves as the agent’s "ears" and "mouth".

Agent Runtime : Node.js execution environment that preserves conversation state, loads plugins, and processes tool‑call requests from the LLM.

OS‑Native Tools : Executes shell commands (e.g., curl, git, npm) and provides a Browser Relay to hijack an already‑open browser for seamless web automation.

Technical Advantages

System Authority : Runs directly on the host with full shell access, enabling real file manipulation, script execution, and code deployment.

Privacy‑First Memory : Uses sqlite-vec together with local Markdown files, keeping data on‑device and supporting hybrid vector‑semantic + keyword retrieval.

Multimodal Interface : Accepts WhatsApp voice (via Whisper), image inputs (Vision models), and outputs text, TTS audio, or file streams.

Hybrid Browser Control : Supports both headless and headed automation, reusing existing Chrome sessions and cookies to bypass anti‑scraping measures.

Applicable Scenarios

24/7 personal assistant – daily briefings, email triage, proactive reminders.

Remote control – mobile commands to manage a home machine, run scripts, or restart servers.

Developer productivity – automated Git workflows, environment setup, and code debugging.

Real‑time monitoring – crypto/finance alerts, market monitoring, and research report generation.

Agent Orchestration (ReAct + Function Calling)

Observe : Gather user input and recent MEMORY context.

Plan/Reason : LLM performs chain‑of‑thought reasoning to decide which tool to invoke.

Act : LLM emits a structured JSON tool call (e.g., read, exec) which the runtime executes.

Reflect : Tool output is fed back to the LLM; if the task is incomplete, the loop repeats, possibly spawning a sub‑agent for long‑running work via the sessions_spawn tool.

Memory Management

Session Context (in‑memory, invisible) – transient state for the current conversation.

Daily Logs ( memory/YYYY‑MM‑DD.md, visible) – raw daily logs stored in the project root.

Curated Memory ( MEMORY.md, visible) – long‑term, user‑editable knowledge base that is periodically compacted.

USER.md – character sheet describing the human user (name, role, preferences).

IDENTITY.md – defines the agent’s persona.

During a session the agent loads USER.md and, in the main session, also loads MEMORY.md. Private data never leaks because these files are only read in direct sessions.

System Prompt Structure

The core system prompt is assembled from modular sections:

Tooling : Lists available tools and brief descriptions.

Skills : Dynamically loads skill definitions when needed.

Memory Recall : Rules for when and how to query MEMORY.md and daily logs.

Self Update : Guidelines for applying configuration changes or updating the agent.

Workspace : Current working directory.

Documentation : Local and remote docs URLs (GitHub, Discord, etc.).

Project Context : Injects the contents of AGENTS.md, SOUL.md, TOOLS.md, IDENTITY.md, USER.md, and HEARTBEAT.md.

Runtime : Host, OS, Node version, model, repository root, and reasoning level.

Prompt modes ( full, minimal, none) control how much of this information is sent to sub‑agents.

Tool List (excerpt)

- read: Read file contents
- write: Create or overwrite files
- exec: Run shell commands (TTY required for interactive CLIs)
- web_search: Search the web via Brave API
- browser: Control a web browser
- sessions_spawn: Spawn a sub‑agent session
- message: Send messages or channel actions
- ... (other tools such as git, cron, canvas, etc.)

Runtime Information

agent=main | host=GEORGEFU-MC1 | repo=/Users/georgefu/clawd | os=Darwin 24.5.0 (arm64) | node=v25.4.0 | model=google/gemini-3-pro-preview | channel=webchat

Safety and Governance

Never exfiltrate private data without explicit permission.

Ask before performing destructive actions (e.g., rm).

Self‑updates ( config.apply, update.run) are only executed when the user explicitly requests them.

In group chats, the agent replies only when directly mentioned, when it can add genuine value, or when correcting misinformation; otherwise it stays silent.

The design emphasizes transparency, user control, and extensibility, making Clawdbot a powerful foundation for building personal AI assistants that operate entirely on the user's machine.