- 27
- April
In late April 2026, the AI Agent Tools landscape gained a new entrant from Thailand — thClaws, an open-source AI Agent Workspace written entirely in Rust. A single 15 MB binary that boots in 200 milliseconds, runs locally on your machine, and ships to market within 3 weeks of a key insight discovered from the Claude Code source code leak. Developed by ThaiGPT Co., Ltd., this article takes a deep dive into what thClaws is, where it came from, what's new, and why it matters.
Quick Summary — What is thClaws?
thClaws is an AI Agent Workspace, open-source and written entirely in Rust — runs locally on your machine (local-first), ships as a single binary at 15 MB, boots in 200ms, supports macOS, Linux, and Windows. It can edit code, automate workflows, search knowledge bases, and orchestrate teams of agents all in one binary. Released under dual MIT/Apache-2.0 license.
The Origin of thClaws — From Source Leak to a 98.4% Insight
thClaws didn't appear by accident — its origin traces back to the Claude Code Source Code Leak in late March 2026, when Anthropic accidentally published a JavaScript sourcemap of Claude Code on an npm package, allowing researchers worldwide to read the internal source code for the first time.
An engineering team at ThaiGPT Co., Ltd. in Thailand was among the groups that read the sourcemap and discovered something that changed how the AI Agent Tools landscape thinks about engineering forever:
The 1.6% Discovery
Actual AI decision logic accounts for only 1.6% of the codebase — the remaining 98.4% is "harness code" — entirely deterministic logic handling permission management, context window, tool routing, sub-agent dispatch, retry policy, and sandbox constraints.
This insight meant — building an AI Agent harness like Claude Code does not require being a frontier AI company. A small team can do it. ThaiGPT therefore decided to rewrite everything in Rust to ship a binary that's small, fast, and portable. It took the team only 3 weeks to ship the first binary in late April 2026.
Technical Specifications
thClaws was designed to be "thin and fast" from day one. Numbers compared to other AI Agent Tools:
| Item | Detail |
|---|---|
| Programming Language | Rust 1.85+ throughout (no TypeScript/JavaScript runtime) |
| Binary Size | ~15 MB (single binary, no Node.js or other runtime needed) |
| Boot Time | ~200 milliseconds (ready immediately after Enter) |
| Supported OS | macOS (Apple Silicon + Intel), Linux (x86_64 + aarch64), Windows (x86_64 + ARM64) |
| Build Prerequisites | Rust 1.85+, Node.js 20+, pnpm 9+ (only for self-build — end users can download the binary directly) |
| License | Dual MIT / Apache-2.0 (open source forever) |
Three Interfaces, One Binary
A standout feature of thClaws is that it ships with 3 interfaces in a single file — pick whichever fits your situation:
| Mode | Best For | Detail |
|---|---|---|
| Desktop GUI | Long sessions / multi-tasking | Native window with Terminal, Chat, Files, and optional Team tabs |
| CLI REPL | Servers, SSH, headless | Terminal-based interactive prompt — works over SSH, no GUI required |
| Non-Interactive | CI/CD, automation, scripts | Single-turn — send prompt in, receive result out, exit. No interactive loop |
Architecture: 100% Local-first
The core design principle of thClaws is "your machine, your agent" — everything stays on the user's own machine:
- Code lives on your filesystem — thClaws reads/writes files in the working directory you specify, no upload to cloud
- Sessions are on disk — the entire conversation history is saved as files on your machine; you can backup, copy, or delete whenever you want
- API keys live in OS keychain — uses macOS Keychain / Linux Secret Service / Windows Credential Manager, not stored as plaintext
- Prompts go directly to the LLM provider — only when calling the LLM (Claude API, OpenAI, Gemini, etc.) does context leave your machine — there's no thClaws middleware server in between
This design makes thClaws ideal for organizations with sensitive data — all code/data stays on the user's machine without going through any thClaws vendor server (because there isn't one).
Multi-Agent Teamwork — Orchestrate a Team of Agents
thClaws supports running as a team of multiple agents via a shared mailbox + task queue:
- Lead Agent — the main agent that takes prompts from the user, breaks tasks into subtasks, and delegates
- Teammates — child agents that pull tasks from the queue, run in parallel, and post results back to the mailbox
- Shared Mailbox — a central channel that all agents share for reading/writing messages between themselves
- Task Queue — a work queue where the lead pushes tasks and teammates pull them when free
This pattern is similar to Claude Team / Multi-Agent Orchestration recommended by Anthropic, but thClaws makes it easier with direct CLI commands and runs everything locally — no separate server to deploy.
Standards Supported — Vendor Neutral
thClaws does not invent new formats — it uses open standards already adopted across the industry, so tooling that works with Claude Code, Cursor, OpenAI Codex, Sourcegraph Amp, and Factory works with thClaws unchanged:
| Standard | Purpose | Adopters |
|---|---|---|
| MCP (Model Context Protocol) | Connect tool servers (GitHub, Slack, Postgres, Filesystem, etc.) | Anthropic, Google, OpenAI, Cursor, Sourcegraph, Factory |
| AGENTS.md | Project-level instructions file that the agent reads before working | OpenAI Codex, Cursor, Sourcegraph, Factory (and now thClaws) |
| SKILL.md (Agent Skills) | Reusable expert workflows that the agent picks automatically | Anthropic Claude.ai / Claude Code / API (and now thClaws) |
Sticking to open standards means users are not locked into a vendor — Skills you write today for Claude Code work tomorrow on thClaws unchanged, and MCP servers you've configured continue to work without rewrites.
Installation
You can install thClaws in two main ways:
- Download a Pre-built Binary — go to thclaws.ai/downloads, choose your OS, and use it directly with no dependencies to install
- Build from Source — clone the repo from github.com/thClaws/thClaws; you'll need Rust 1.85+, Node.js 20+, pnpm 9+, then run
cargo build --release
After installation, run a single command:
# Open the Desktop GUI
thclaws
# Open CLI REPL in terminal
thclaws --cli
# Non-interactive (single turn)
thclaws --print "Explain the bug in main.rs"License + Open Source
thClaws is released under a dual MIT / Apache-2.0 license, the standard open-source pairing in the Rust ecosystem. Users can:
- Use it freely without restriction (including commercial use)
- Modify the source code, fork, and redistribute
- Embed it inside their own products (both open-source and proprietary)
- The development team has stated explicitly that the client will be free and open source forever
thClaws vs Claude Code — Side-by-Side
| Item | thClaws | Claude Code |
|---|---|---|
| Language | Rust | TypeScript / Node.js |
| Install Size | ~15 MB (single binary) | ~200+ MB (Node.js runtime required) |
| License | MIT / Apache-2.0 (open-source) | Proprietary (Anthropic Commercial Terms) |
| LLM Provider | Vendor-agnostic (Claude, GPT, Gemini, etc.) | Anthropic Claude only |
| GUI | Native Desktop GUI built-in | CLI / IDE Extension only |
| Standards | MCP + AGENTS.md + SKILL.md | MCP + CLAUDE.md + SKILL.md |
Summary — Why thClaws Matters
| Topic | Summary |
|---|---|
| What it is | An open-source AI Agent Workspace written entirely in Rust |
| Who built it | ThaiGPT Co., Ltd. (Thailand) |
| Origin | Discovered from the Claude Code source leak that 98.4% is harness code → rebuilt from scratch in Rust within 3 weeks |
| What's new | Single 15 MB binary, Local-first, three interfaces in one binary, Multi-agent teamwork, Vendor-neutral |
| License | Dual MIT / Apache-2.0 — free and open source forever |
thClaws proves that frontier-class AI Agent Tools don't have to come from giant companies — a small Thai team built one in 3 weeks once they understood what's "harness" and what's "intelligence".
- Saeree ERP Team
Read More — Related Articles
- Claude MCP: What is Model Context Protocol — A Complete Deep Dive
- Claude Skills: What are Agent Skills — A Complete Guide
- Source Leak: Claude Code Source Code Leak: 512,000 Lines
