thClaws

New in v0.126.0 — many agents, one workspace

A workspace is no longer one agent. It is a shelf of them. Keep your coding agent, a research agent, and a book-writing agent side by side in the same project, and switch between them from a rail on the left without losing anything — each one keeps its own conversation, scroll position, and half-typed message while you are away.

┌────┬───────────────────────────────────────────────┐
│ MA │  Chat · Terminal · Files · UI                 │
│ RE │                                               │
│ BA │  every agent: its own sessions, memory,       │
│    │  settings, and browser logins                 │
│  + │                                               │
└────┴───────────────────────────────────────────────┘

• Each agent is self-contained. Its own sessions, knowledge bases, settings, and browser profile, in its own folder under .thclaws/bots//. An agent's file tools stop at that folder, so one agent cannot read or write another's work.
• They run side by side. Every agent is its own process, supervised by the app. An agent that crashes is restarted without disturbing the others, and an agent you switch away from keeps working.
• Add an agent from the rail's +. Get any Agent Template from thClaws.cloud, or start a blank agent that behaves exactly like opening a new folder. Up to 8 agents run at once.
• Desktop and browser alike. The same rail works in the desktop app and in thclaws --serve.

Your existing workspaces upgrade themselves. The first time v0.126.0 opens a workspace — in the desktop app or with --serve — it moves your project into .thclaws/bots/main/ (git history included) and tells you it did. Scripts keep working: thclaws -p and --cli at the workspace root step into that first agent automatically. An older thClaws opening an upgraded folder finds a note asking you to update, rather than a broken project.

Before you upgrade: anything that holds your project's old path — an editor window, a script with an absolute path, another clone — needs the new one. Commit or back up first if that matters to you. Your home directory is never upgraded, and hosted workspaces on thClaws.cloud are not migrated by this release.

Full list of changes: CHANGELOG.md

See it work

Three tabs, one binary — captured from a live thClaws session looking at its own source.

Highlights

Media Studio — generate images & video, built in

A point-and-click shell for the new media tools — Text → Image, Image Edit, Text → Video, and Image → Video — across Google Gemini, OpenAI gpt-image-2, Alibaba Qwen, Veo, and DashScope HappyHorse. Pick a provider/model, set the resolution, hit Generate; finished media lands in a live gallery you can click to reuse as the source for the next run. The same tools (TextToImage / ImageToImage / TextToVideo / ImageToVideo / MediaJobStatus) are callable by the agent from chat too — async video included.

OpenRouter Fusion — many models, one answer

openrouter/fusion fans your question out to a panel of up to 8 models in parallel (each can search the web), then a judge weighs the answers and synthesizes the consensus — multi-model deliberation packaged as a single model id. The new openrouter/fusion+ makes it fully configurable right from the model picker: choose the panel, the judge, the per-model tool-call budget, temperature, reasoning effort, and tool-choice. A practical stand-in when the model you'd reach for isn't available on the platform.

Docs: media tools · Media Studio · Fusion setup

Recently shipped

A release most weeks. Full history in the changelog; the things most likely to matter to you:

• Enterprise policy + client-side audit trail — a signed org policy file turns the same binary into a managed client: branding, an extension allow-list, a forced LLM gateway, OIDC SSO, a forced permission mode, and a payload-free record of every tool call. (ENTERPRISE.md · ภาษาไทย)
• OS-level Bash confinement, on by default — shell commands run inside a kernel-enforced filesystem boundary (macOS Seatbelt, Linux Landlock, bubblewrap fallback), so an obfuscated command still can't write outside the workspace.
• Thai PII masking — an opt-in pipeline that detects and tokenizes Thai personal data before it reaches the model, then restores it in the output.
• Movie Maker — screenplay → shots → generated video, end to end.
• Browser automation — drive a real Chrome session as a tool.
• Workflows — deterministic multi-agent scripts (/workflow run).
• GUI Shells — ship a custom HTML surface with your agent, themed and bridged to sessions, memory and tools.
• thclaws.cloud — an agent catalog, a hosted runtime, and a metered gateway for people who don't want to hold provider keys.

Hacking in public

thClaws started in April 2026. Since then it has shipped 100+ public releases, drawn 30+ contributors, and lands roughly a release a week — the current release is v0.121.0. It's developed by a small team at ThaiGPT Co., Ltd. — and a meaningful chunk of the codebase comes from outside contributors who heard about it and stayed.

We're aiming for v1.0 = "the multi-platform agent": the same agent loop on your desktop, in your terminal, and bridged into Telegram, Discord, Slack, WhatsApp, Facebook Messenger, and LINE. Telegram, LINE, and Messenger are already shipping. Discord, Slack, and WhatsApp are next — and they're great places to plug in. (Contribute →)

Built in Thailand. Meant for the world.

Four surfaces, one engine

The same Agent loop, Session, and ToolRegistry back every UX:

• Desktop GUI (thclaws) — native window with Terminal, Chat, Files, and optional Team tabs.
• CLI REPL (thclaws --cli) — interactive terminal prompt for SSH, headless servers, or zero-GUI workflows.
• Non-interactive mode (thclaws -p "prompt") — single turn, exits. Pipe-friendly for scripts and CI. -v for token usage on stderr.
• Webapp (thclaws --serve --port 7878) — same engine over WebSocket/HTTP, plus a drop-in API surface so other tools can drive the agent: OpenAI-compatible at /v1/chat/completions and Anthropic-compatible at /v1/messages. Point the openai or anthropic SDK at it, or set ANTHROPIC_BASE_URL. SSH-tunnel for "Claude Code anywhere" without opening a port.

Features

Everything's in one binary. Pick the surface that fits the task, swap the provider, drop in a skill, glue in an MCP server, then walk away while a scheduled job or background agent finishes the work.

• Multi-provider — Anthropic (native + Claude Agent SDK via Claude Code auth), OpenAI (Chat Completions + Responses/Codex), Google Gemini & Gemma, Alibaba DashScope (Qwen), DeepSeek, Z.ai (GLM Coding Plan), NVIDIA NIM, NSTDA Thai LLM (OpenThaiGPT, Typhoon, Pathumma, THaLLE), xAI (Grok), Moonshot (Kimi), MiniMax, OpenRouter, Agentic Press, Azure AI Foundry, Ollama (local + Anthropic-compatible + Cloud), LMStudio, plus a generic OpenAI-compatible slot (oai/*) and first-class LiteLLM for internal proxies. The shipped catalogue carries 1,100+ models across 27 providers — LLM plus image, video and audio. Switch mid-session with /model or /provider.
• Open standards, not a walled garden — Model Context Protocol for tools, AGENTS.md for project instructions (adopted by Google, OpenAI, Factory, Sourcegraph, Cursor), SKILL.md with YAML frontmatter for packaged workflows. Configuration portable between thClaws, other compliant agents, and whatever comes next.
• Skills, plugins, MCP servers, hooks — extend the agent without touching Rust. Skills are folders with a SKILL.md. Plugins bundle skills + commands + agent definitions + MCP servers under one manifest. MCP brings in third-party tools (GitHub, filesystems, browsers, Slack…) over stdio or HTTP-Streamable with OAuth 2.1+PKCE. Hooks run shell scripts on lifecycle events (pre_tool_use, permission_denied, session_start, …).
• Three tiers of agent orchestration — model-driven subagents (Task tool, blocking, up to 3 levels deep); user-driven concurrent side-channels (/agent , parallel to main, own cancel token); multi-process Agent Teams with shared mailbox, task queue, tmux panes, and optional git worktrees.
• Knowledge bases (KMS) + /dream — per-project and per-user wikis under .thclaws/kms//pages/, indexed by a one-line index.md. Grep + read (no embeddings), following Andrej Karpathy's LLM-wiki pattern. /dream mines your recent sessions in the background and writes a dated audit-trail page to review with git diff.
• Plan mode — EnterPlanMode proposes an ordered list of steps you Approve / Cancel / Skip / Retry. Same UX in GUI sidebar and /plan slash command.
• Schedule recurring jobs — /schedule add runs an agent on cron, fixed intervals, or filesystem changes (watchWorkspace). In-process scheduler for ephemeral, native daemon (launchd / systemd-user) for survives-reboot.
• Long-running loops & overnight builds — /loop for fixed-interval iteration, /goal for audit-driven completion. /goal --auto is a Ralph-style overnight builder that keeps going until the goal is satisfied or you wake up.
• Document workflow — native PDF, DOCX, PPTX, XLSX read + edit + create tools, plus image rendering. Ingest a 50-page PDF, summarize into KMS, produce a follow-up deck — one conversation.
• Memory & project instructions — AGENTS.md (or CLAUDE.md) walked up from cwd and injected into the system prompt. Persistent memory store classified as user / feedback / project / reference, stored as markdown you can read, edit, or commit.
• Settings as one file — .thclaws/settings.json (project) or ~/.config/thclaws/settings.json (user). API keys go in the OS keychain by default (macOS Keychain / Windows Credential Manager / Linux Secret Service) with .env fallback for CI.
• Session resume — thclaws --resume last or --resume . Sessions live as JSONL under .thclaws/state/sessions/ — git-friendly, grep-friendly, never opaque.
• Safety first — shell commands run inside an OS-enforced filesystem boundary (macOS Seatbelt, Linux Landlock, bubblewrap fallback), so obfuscation doesn't get a command out of the workspace; modes workspace (default) / strict / off. On top of that: a filesystem sandbox scoped to the working directory, destructive commands flagged, and your approval on every mutating tool call unless you've opted into auto-approve. Permission requests label which agent is asking when multiple are running.
• Offline-capable — Ollama (native + Anthropic-compatible) lets you run entirely against a local model. No cloud round-trip, no API key.
• Media, end to end — Media Studio generates images and video (Gemini, OpenAI gpt-image, Qwen, Veo, LTX, iApp for Thai text) from a point-and-click shell or from chat; Movie Maker takes a screenplay through shots to a finished film.
• Browser automation — drive a real Chrome session as a tool: navigate, read the page, fill forms, capture screenshots.
• Workflows & GUI Shells — /workflow run scripts deterministic multi-agent pipelines; GUI Shells let you ship a custom HTML surface bridged to sessions, memory and tools, themed with the app.
• Thai PII masking (opt-in) — detects and tokenizes Thai personal data before the model ever sees it, then restores it on the way out. Per-workspace toggle.
• **Enterprise Edition — same bi