Truss

Local-first coding agents<br>A coding agent that works your way.<br>Plan, edit, and review code with local models from the shell, terminal, VS Code, or desktop—with the same tools, workspace state, and approval controls everywhere.<br>Download desktopGet started

truss-harness CLIlocal workspace<br>$ truss-harness chat "Review the current diff"▋↵<br>model: qwen3-coder | mode: edit | permissions: ask<br>assistant I found three changed files. I will inspect the diff first.<br>tool git diff --stat<br>ready for local execution

One runtime, four ways to work<br>Bring the agent to your workflow.

Start in a shell, stay in your terminal, work inside VS Code, or open a dedicated desktop workspace without changing your local model setup or safety controls.<br>Meet the clients<br>Client surfaces<br>Choose the interface, not a different agent.<br>One agent across your shell, terminal, editor, and desktop workspace.

CLIAutomate from your shell<br>Run coding tasks, discover local models, and reuse workspace profiles from scripts or your normal terminal.<br>Explore CLI<br>TUIKeep your flow in the terminal<br>Use a full-screen workspace with files, editor, Git diff preview, chat, shell output, approvals, and model controls.<br>Explore TUI<br>VS CODEWork where you edit<br>Get streaming chat, file context, inline completions, agent modes, approvals, and Git commit-message help in VS Code.<br>Explore VS Code<br>DESKTOPOpen a dedicated workspace<br>Use the standalone Electron app for file browsing, diffs, terminal commands, chat, approvals, and a built-in Git workflow.<br>Explore desktop

01Local by default<br>Discover and connect Ollama, LM Studio, llama.cpp, or another OpenAI-compatible local endpoint—without a cloud account.

02Cloud on your terms<br>Add a supported provider through an API key you control. Truss does not require a hosted control plane or proxy model requests.

03State that carries forward<br>Conversations, implementation plans, workspace memory, Git-aware status, and deterministic commands keep work moving.

04Context with intent<br>Open files, recent edits, Git changes, plans, and workspace memory give the agent useful context without sending the whole repository.

05Approvals stay visible<br>Choose the permission policy that fits the task, then review requested reads, writes, terminal commands, and network work before it runs.

06Built to be replaced<br>Providers, tools, context sources, memory, and clients are separate interfaces—so new capabilities do not rewrite the agent runtime.

Stay connectedFollow the project and see what ships next.<br>GitHub ↗LinkedIn ↗

Built for thoughtful software work.© 2026 Truss