Truss for VS Code

Truss is a local-first coding-agent side panel for VS Code. It connects to Ollama, LM Studio, llama.cpp server, and custom OpenAI-compatible local endpoints through the Truss runtime service. It does not require a cloud provider account.

Install

Install Truss for VS Code from the Marketplace, then open the Truss icon in the Activity Bar. The extension includes its runtime service, so the CLI is not required.

In Settings, select a detected local server or enter an endpoint, refresh models, and set the context window, permission policy, and optional internet research access. Use the bottom control bar to switch models and agent modes during a session. Agent responses render Markdown and formatted code blocks. Type @/ in the chat composer to fuzzy-search workspace files; choose a file with arrow keys and Enter or Tab to attach its bounded contents to the next prompt. Slash commands such as /status, /update, and /help remain reserved for local workspace actions.

To add local MCP tools, enter an mcpServers JSON object in the MCP settings field. The extension passes it to its runtime service; Agent mode loads enabled servers and Plan mode loads only servers marked readOnly. Connections restart after settings change and their status is shown in the panel.

Side panel

The side panel contains:

• Conversation history with per-conversation delete controls and live streaming chat
• Local server, endpoint, context-window, and permission controls
• Bottom controls for Chat, Plan, and Agent modes plus the active local-model dropdown
• Tool approval prompts
• A stop-generation control
• Source Control title-bar action that generates a commit message from staged or unstaged Git changes and fills Git's commit-message input
• A Help button for deterministic workspace commands

Conversation transcripts are stored in VS Code workspace state. They remain available after closing or reloading the panel and across extension restarts. Use the x control beside a conversation to delete it. The live runtime session itself is recreated on demand from the saved user/assistant transcript, so provider state is not treated as durable storage.

Changing the bottom model dropdown restarts the Truss runtime with the selected model. For Ollama endpoints, Truss also makes a best-effort request to release the previous model from memory; compatible servers select the new model on the next request according to their own server lifecycle.

Modes govern the agent's available tools:

Permissions apply after the selected mode:

Internet research adds bounded web_search and web_fetch tools. It is disabled by default and still prompts under Ask and Auto-allow read-only.

Workspace commands

Type these directly in the chat composer. They are executed locally, not sent to the model, so they also work without a selected model.

The Command Palette exposes the same maintenance actions as Truss: Initialize Workspace Instructions, Update Workspace Memory, Show Workspace Status, and Clear Workspace Memory.

Inline completions

Truss registers VS Code inline completions for editor documents. When a completion is available, accept it with the normal editor inline-completion binding, usually Tab. Keep a small local model responsive by using a dedicated coding model and a local server with native tool calling enabled.

Settings

Package for the Marketplace

The extension is bundled before packaging so its runtime dependencies are included. From this package directory, run:

npm run package

This produces a .vsix file for local installation or Marketplace submission. Set the final publisher, name, version, repository links, and a PNG Marketplace icon before publishing under an organization account.