The coding agent that runs on your machine.

1Install

Requires VS Code 1.120 or later.

• Extensions view (Ctrl+Shift+X) → search GBS SE Agent → Install
• Or run code --install-extension gbs-internal-toolchain.gbs-software-agent

2Choose a model

Ollama — local, recommended

Install Ollama, then pull a model that supports tool calling:

ollama pull qwen3-coder:30b     # capable GPU (~20 GB)
ollama pull qwen2.5-coder:14b   # mid-range GPU
ollama pull qwen2.5-coder:7b    # most laptops

Open the chat panel, click the model chip under the input and pick Ollama and your model. Models ending in :cloud run on Ollama's servers, not yours — the chip marks those.

LM Studio, llama.cpp, vLLM

Any OpenAI-compatible /v1 endpoint works, and needs no API key on localhost.

Set gbsAgent.openai.baseUrl in your User settings, choose OpenAI-compatible in the model chip, and set gbsAgent.contextWindow to the context length the server was started with.

Cloud

Set gbsAgent.openai.baseUrl to your provider, run GBS Agent: Set API Key, then enter the model id. The key lives in VS Code's secret storage and is only sent to that endpoint. Before any code leaves your machine you're asked once per destination.

3Run it

1. Open your project and trust the folder — the agent stays off in Restricted Mode.
2. Click the robot icon in the activity bar.
3. Describe the task and press Enter.

Add rate limiting to the login endpoint, 5 attempts per minute per IP, and make sure the auth tests pass.

• The file you have open, and anything you've selected, is attached automatically.
• Keep typing while it works to steer it; Esc stops it.
• Edits apply as it goes. Shell commands, and edits to config, scripts or credentials, ask first.
• Each turn ends with the files changed, a diff button, and what it cost in tokens.
• Watch the context meter next to the send button; /compact frees space, /new starts over.

4Teach it your stack

Short, always-true facts — build and test commands, conventions — belong in AGENTS.md at the project root. Bigger knowledge that only matters sometimes belongs in a skill: only its one-line description stays in the prompt, and the instructions load when the task matches.

.gbs/skills/component-library/
├── SKILL.md            # the rules that matter
└── references/forms.md # detail, loaded only when needed

---
name: component-library
description: Builds UI with the GBS component library. Use when changing UI in apps that import @gbs/ui.
autoAttach: ["src/**/*.tsx"]
---

# Component library
- Tables: `DataGrid`, never a raw table element.
- Import from `@gbs/ui`, never from `@gbs/ui/dist/*`.

Skills can live in the project, ship inside a package your team installs, or sit in a personal folder used across every project. Run GBS Agent: New Skill to scaffold one, or type /skills to see what's loaded.