CAGE — Composable Agent Graph Engine

A self-hosted AI coding assistant. Your models, your hardware, your code — nothing leaves the machines you control unless you point a graph at a remote endpoint yourself.

This extension is the editor client. It connects to a CAGE server that you run: the server owns model loading, GPU placement, the agent loop, retrieval and the multimodal pipelines; the extension renders the conversation and executes tool calls locally in your workspace.

You need a CAGE server to use this. The extension does nothing on its own. See Requirements.

The CAGE chat panel in VSCode

What makes it different

The agent loop is a graph you can edit, not fixed code. Prompt nodes, format checks, branches, loops, and nodes that pause to ask you a question — laid out visually. Every node picks its own model, quantization, GPU set and sampling, so a small fast model can plan while a large one writes the code, in one run.

The Logic Editor

Files are approved one at a time, as they are written. Each write_file or apply_patch shows you that file's diff and waits. Approve it and it lands in your workspace; decline and the model is told so. Only one question waits at a time.

Tools run in your editor, not on the server. Reading and writing files, applying patches, terminal commands, builds and tests all execute locally, with per-tool approval and risk levels. MCP servers can be attached and their tools are offered alongside the built-in ones.

Retrieval over your project. The client pushes a full index and then incremental deltas as you edit; chunking is tree-sitter AST-aware. Only a file listing crosses the wire until content is actually needed.

Project memory. The server records what changed in each file and why, and brings that back when the file is in play again. You can add your own standing rules and say when they apply, in plain language.

Multi-GPU placement that reads your machine. Placement is decided from the real topology — NVLink cliques, free VRAM per card, dense vs mixture-of-experts. You pick the GPUs; split mode, layer counts and expert CPU offload are worked out for you.

Requirements


A CAGE server	Linux x86-64, plus a backend pack for your GPU. NVIDIA, AMD, Intel and CPU-only are all supported
VSCode	1.85 or newer
Models	Yours. The server ships no weights

Install the server first — the extension has nothing to connect to otherwise. Server installation guide →

Getting started

Install and start a CAGE server.
Install this extension.
Set cage.server.host and cage.server.port to point at it.
Open the CAGE panel, pick a graph in the selector at the bottom, and send a message.

Full walkthrough: Extension installation guide →

Documentation


Client guide	Every panel, every node type, every setting
Server guide	The three server processes, models, backends, retrieval, accounts
What's new	Changes in this release, and what needs action when upgrading

Licence

The extension is free to install. The CAGE server is a commercial product, sold as a base product plus separately purchased modules, each with perpetual updates. Free evaluation builds are available for testing and may not be used in production.

Full terms: LICENSE. Third-party components and their licences: THIRD-PARTY-NOTICES.md.