The ForgeFor this device
TALOS Desktop
The full workshop.
Everything a developer needs — and a researcher: agents as a graph, context handled for you, memory, deep research and local models, on Windows.
TALOS / Products
PlateII
Build2026-10-09 · 1adb128
The Hall of Steles · Three surfaces
One forge. Three works.
One environment, your models: on Windows, in the terminal, on your phone.
The Steles
Each with its own craft.
The AnvilFor this device
TALOS CLI
Fast. Lean. Just code.
A coding agent in your terminal, on the model you choose: every tool asks first, every change can be undone, every session can be resumed.
The WatchtowerFor this device
TALOS for Android
An assistant that runs your phone.
Say “Hey TALOS”: it sees what is on the screen and acts across your phone, with a model on the device or the cloud model you choose.
The Measure
Measured side by side.
From each product's README and release.
| TALOS Desktop | TALOS CLI | TALOS for Android | |
|---|---|---|---|
| Platform | Windows 10 1809 or later, x64, or Windows 11 x64 | Node.js 24 (24.18 or later in the 24 line). Tested on Windows 11 x64 | Android 8.0 or later (API 26), 64-bit ARM |
| Status & version | Public release · v0.1.25 | npm · v0.5.2 | Testing build · v0.1.40 |
| Install | Installer or portable ZIP. No administrator rights | npm i -g talos-code, then talos | APK from the TALOS releases on GitHub |
| Local models | llama.cpp engine included, CPU and Vulkan; GGUF models downloaded in the app, checksum verified; Ollama and LM Studio too | Ollama, LM Studio and llama.cpp, with no key | GGUF models through llama.cpp on the device; Hugging Face browsing with a fits-in-memory check |
| Cloud providers | 23 providers, from OpenAI and Anthropic to Bedrock and Vertex AI, with your own key | 22 providers, from OpenAI and Anthropic to Bedrock and Vertex AI | OpenAI, Anthropic, Gemini, DeepSeek, OpenRouter, and Ollama by endpoint |
| Memory | Memory, with notes and tasks, kept by the workspace | — (sessions can be resumed and forked) | Typed personal and project memory; every write is a request you approve |
| Agents | Agents and verifiers, followed live as a graph | One coding agent; each tool allowed, asked or denied | One agent with 69 typed tools, each allowed, asked or denied, plus Codice, a coding agent with sub-agents |
| Research & library | Library and Deep Research | — | Web search, durable research runs and an on-device Library |
| Voice & phone control | — | — | “Hey TALOS” wake word, dictation, replies in your own voice; acts on the phone and inside other apps, optional ADB bridge |
| Works offline | With a local model: the network is used only for remote providers, web search and model downloads | With a local engine | With a local model: inference, wake word and your data stay on the device |
| Telemetry | None | None | No TALOS backend: conversations, memory and settings stay on the device |
Sources
- TALOS Desktop for Windows — README
- TALOS Harness Desktop — README
- TALOS Desktop — Model Lab, provider list (checked 2026-09-25)
- talos-code — README on npm
- TALOS for Android — README
- TALOS releases on GitHub
A dash means the product's documentation for that version does not state it — not that it never will. Versions and requirements are read from the releases at every build.