Portable Windows desktop ·

The portable harness
for serious developers.

Bring your own model. Plug in the subscriptions you already pay for, or skip subscriptions and run fully local.

Download for Windows See how it works

Windows 10 / 11 · Single executable · Free to download

The smartest model changes every week.
The harness you drive it with shouldn’t.

What it's for

Built for the work you actually do.

Five jobs we obsess over. Pick the one that's yours; the rest will still be useful tomorrow.

For Game Developers

Built into the engine, not bolted onto it.

Native plugins drive Unity, Unreal Engine, Godot, and Blender from one prompt — scenes, materials, prefabs, Blueprints, physics — then batch the work into a single undo group. Three.js, React Three Fiber, and WebGL workflows are backed by version-pinned PBR, animation, and post-processing knowledge. And they're quick to connect: just ask the agent and it installs the bridge for you. No MCP setup, no JSON, no plugin to download.

  • Four quick-connect engine plugins. Unity, Unreal (5.7 & 5.8), Godot, and Blender — just ask and the agent installs the bridge.
  • Three.js / R3F / drei / Rapier patterns baked in. Color-managed by default. No washed-out materials.
  • Action Capture verification. Event-synced screenshots prove a transient action happened before claiming done.
  • Decisions at game speed with TypeSafe AI. NPC, enemy, and simulation choices come back as typed answers with confidence in a few hundred milliseconds, and the agent wires them into Unity, Godot, Unreal, or the browser from one prompt.
Explore the plugins

For Prototypes

Idea Monday. Working demo Tuesday.

Go from a sentence to a running app without picking a stack first. Scaffold the project, generate placeholder assets, embed the live preview, and ship a one-page spec alongside the code.

  • Scaffold from a sentence. Move from spec to file structure, working dependencies, and first run.
  • Image generation, masked edits, and reference-based variants without leaving the chat.
  • File history reverts any edit. Every command needs your permission. Brakes are built in.

For New Codebases

Walk into 200,000 lines. Be useful by lunch.

Symbol-aware search and reference lookup finds the file you'd have grepped for in one prompt instead of ten. A read-only Code Explorer task agent handles the legwork so your senior model stays on architecture.

  • Tatsu's codebase index. References, symbols, and project fingerprint without configuring a thing.
  • Code Explorer task agent. Cheap model reads, expensive model decides.
  • AGENTS.md and Layered Memory keep your project context across sessions and providers.

For the Web

Edit. See. Edit. See. Without alt-tabbing.

DevBrowser is a real headed Chromium the agent can drive. Screenshots, console logs, network, DOM, and storage are all reachable in the same conversation. Tweak CSS, reload, capture, confirm. The browser is Tatsu's own, opened alongside the editor and isolated from your personal Chrome.

  • DevBrowser. Real Chromium with DevTools, cookies, storage, and network interception.
  • Visual verification loop. Fix, reload, screenshot, diff, confirm.
  • Local static server included. A Tatsu-owned Chromium window keeps cookies, profile data, and extensions separate from your system browser.

For the Portable & Private

One signed .exe. No installer. Nothing leaves unless you say.

A single Windows binary is digitally signed, IT-approvable, and runs from a USB drive on a locked-down corporate laptop. No admin rights, no installer, no registry footprint. Fully offline with the local model stack you already use. Credentials stay on disk in %UserProfile%\.tatsu\, the same place every other CLI keeps its keys, ACL'd to your Windows account.

  • Digitally signed by Studio Tatsu via Azure Trusted Signing. Passes Windows SmartScreen on first launch with no "unknown publisher" prompt and no false-positive AV flags. Whitelist by certificate and IT is done.
  • Single portable executable. Runs from a USB stick on an air-gapped workstation. No admin rights, no installer, no registry writes.
  • Fully offline with llama.cpp. Point Tatsu at a llama.cpp server and pull Gemma, Qwen, Kimi, DeepSeek, or any GGUF you've already downloaded — or use Ollama or LM Studio if you prefer. Code with a coding agent for $0 a month, with no subscription anywhere in the chain.
  • Zero telemetry from the desktop app. Credentials, sessions, and history all stay on disk.

Out of the Box

All the tools. All the models. Zero setup.

Every other harness ships an empty shell. You wire up MCP servers, install plugins, and beg for permissions. Tatsu Code ships ready, and every model in the picker can use everything in the box, including the model that's best at the job.

No MCP. No plugins. No JSON.

Tell the agent “connect to Unreal” and it does — it finds the editor, installs the bridge, and connects. Ask for a PowerPoint deck and it writes one. Need a screenshot of your localhost? It captures one. Every other harness ships empty. We ship loaded.

  • Unity, Unreal, Godot & Blender
  • Word, Excel, PowerPoint
  • DevBrowser (real Chromium)
  • PDF analysis with vision
  • Image generation + editing
  • Desktop automation
  • YouTube transcript + video
  • Web search
  • TypeSafe AI decision models
  • Cross-Talk between windows

The right model for the right job.

Different tasks demand different models. The YouTube Agent runs on Gemini because it is the frontier model that actually watches the video instead of just reading the transcript. The Code Explorer runs cheap, so your senior model stays on architecture. You don't pick. The harness does.

  • YouTube AgentGemini
  • PDF ReaderVision-capable
  • Code ExplorerCheap + fast
  • Image GeneratorImage-capable
  • Typed decisionsTypeSafe AI
  • Main chatYour pick

Almost every model on Earth, in one picker.

OpenRouter routes you to 200+ models: every frontier closed model, every open-source SOTA, every image generator, plus the unannounced “stealth” previews labs ship before public reveal. Combined with your direct subscriptions and local models, the picker is a complete catalog of the AI industry.

  • 200+ models via OpenRouter
  • Stealth preview models
  • Image generation models
  • Direct Claude / ChatGPT
  • Local llama.cpp / Ollama / LM Studio
  • One picker. Switch any time.

Bring Your Own Model

You already pay for the model.
Stop paying for the tool.

Tatsu Code is a free harness. You bring the model access, and we never sit between you and your provider.

Claude Pro Claude Max ChatGPT Plus ChatGPT Pro OpenRouter llama.cpp Ollama LM Studio TypeSafe AI Gemini CLI · soon

Plug in your Claude subscription.

Tatsu Code spawns Anthropic's claude CLI locally. Your tokens live in ~/.claude/, where Claude Code put them. We never see them.

Plug in your ChatGPT subscription.

Sign in via OAuth directly inside Tatsu Code with no extra CLI to install. Tokens stay local in %UserProfile%\.tatsu\credentials.enc, encrypted at rest. Your bill stays with the vendor.

Plug in zero.

Run llama.cpp on your laptop. Code with a coding agent for $0 a month. No vendor in the loop.

Tatsu Code is a free desktop harness; you supply the model access. We never resell, pool, charge for, or proxy vendor capacity. There is no Tatsu Code server. The same goes for TypeSafe AI, integrated as a decision-model provider under your own key.

Cross-Agent Relay

Tag-team Claude, Codex, Gemini,
and your local model on the same task.

One agent's run becomes the next agent's starting point. Shared chat history across every provider; durable on-disk archives from every CLI run. No vendor in the middle.

Claude Pro

Diagnose

report.md

Codex (OpenAI)

Fix

session

Gemini

Review

session

llama.cpp (local)

Test

Run Claude Pro on the diagnosis. Hand the result to Codex for the fix. Bring Gemini in for review. Have your local model write the tests. All in one Tatsu Code session. The chat history travels with you across every provider, so the next agent starts with full context. CLI-driven runs (today, that's Claude) also emit a structured report.md archive on disk as the durable, vendor-neutral half of the handoff. Either way: no re-explaining, no copy-paste, no vendor cooperation required.

No platform vendor will ever ship this. Anthropic isn't going to recommend OpenAI; OpenAI isn't going to recommend Anthropic. Tatsu Code can. That's the whole point.

Cross-Talk

Windows that talk to each other.
Whatever model each one runs.

Inside one window, your main model already hands work to task agents and other providers. Cross-Talk is the second kind: open a second Tatsu Code window, on any provider, and the two windows can message each other. Same machine, no cloud in between, and the receiving window's model acts within its own permissions.

Gemini window

Asks

message

Claude Code window

Runs the tests

reply

Gemini window

Reads the report

Every window has an address. Say “send the other window a summary of what you changed” and the agent lists who is reachable, sends, and shows a receipt. On the other side the message arrives under a distinct Message from block, never as something you typed. An idle window starts a turn at once; a busy one queues the message for its next safe point.

Claude Code windows keep their native session messaging for other Claude sessions and use the same bus for everything else, so a Gemini window and a Claude window are one message apart.

How Cross-Talk works

FAQ

Questions that come up.

Yes. It is free to download today, with no paid tier and no telemetry from the app. The version you download is yours to keep, regardless of how pricing or ownership might evolve later. If that ever changes, whether through a paid tier to fund development, a Studio license, or an acquisition, we'll be transparent about it here and in the release notes.
Locally, in %UserProfile%\.tatsu\credentials.enc. They are encrypted at rest and ACL'd to your Windows account. They are never sent to any Tatsu Code server; there isn't one. Tatsu Code is a desktop app, so the only network calls are from your machine straight to whichever provider you've configured.
Yes. Claude routes through Anthropic's claude CLI installed on your machine. Your tokens live in ~/.claude/ where Claude Code itself stored them, and we never see them. ChatGPT (Codex) integrates via OAuth directly inside Tatsu Code for tighter harness compatibility. There is no extra CLI to install; tokens stay local under %UserProfile%\.tatsu\, ACL'd to your Windows account, and never touch any Tatsu Code server (there isn't one). Gemini CLI subscription support is on the roadmap; in the meantime, Gemini models are reachable through OpenRouter.
Yes. The executable is digitally signed by Studio Tatsu via Azure Trusted Signing, a Microsoft-trusted certificate authority. It passes Windows SmartScreen on first launch, plays nice with AppLocker / WDAC by certificate, and avoids the false-positive antivirus flags that catch unsigned binaries. No installer, no admin rights, no registry writes. IT can whitelist by publisher and forget about it.
No, if you're running fully local with llama.cpp (or Ollama / LM Studio). Pull Gemma, Qwen, Kimi, DeepSeek, or any GGUF you've already got. Yes, if you're using a hosted provider. That is the same connection you'd need for that provider's own tool.
Yes, two ways. Inside one window, the main model already delegates to task agents and hands sessions across providers with Cross-Agent Relay. Between windows, every Tatsu Code window on the machine has an address, and its model can list the others and send them messages, whatever provider each runs. The message shows up in the other window as a distinct Message from block and that window's model acts within its own permissions. It all stays on the machine: loopback only, with a token per window. See Cross-Talk.
No. The desktop binary is closed-source. The bug tracker and feature requests are public on GitHub at TatsuAI/TatsuCode.
Currently, our tools and pipeline are focused on Windows, supporting PC and console game development and native Windows desktop applications. At this time, there are no plans to support other platforms. If there is significantly high demand and adequate funding, we will consider including Mac and Linux builds in the future.
Studio Tatsu, LLC. We're a small studio shipping out of the gamedev industry. Not VC-funded. We made it because we wanted it.