Truss

Truss is a local-first, provider-neutral coding-agent platform. Use the same
models, tools, permissions, workspace memory, and plans from the command line,
terminal, VS Code, or a dedicated desktop workspace.

Create multiple workspace-local managed-agent profiles when a task benefits
from separate models or providers. Chat and Plan agents can work concurrently;
Edit agents are serialized through a workspace write lease so they cannot
mutate the same workspace at the same time.

Bring an Ollama, LM Studio, llama.cpp, or compatible endpoint—or configure a
supported cloud provider with your own API key. Truss does not require a hosted
control plane.

Choose a client

Managed agents

Desktop, VS Code, the CLI, and a host-authorized Truss Go connection can
control independent workspace-local agent profiles. Each profile has its own
provider, endpoint, model, mode, and approval policy while credentials remain
with the host. Read the managed-agent guide
for run policy, safety behavior, and CLI commands.

CLI

Use the CLI for direct tasks, scripts, and automations.

npm install --global @truss-harness/cli
truss-cli models
truss-cli chat "Review the current diff"

Terminal UI

The terminal UI combines chat, files, editing, Git diffs, shell output, model
controls, and approvals in one keyboard-first workspace.

npm install --global @truss-harness/tui
truss-tui

VS Code

Install the Truss VS Code extension, open the Truss Activity Bar view, then select a detected local server or enter your endpoint and model. The extension supports streaming chat, focused file context, inline completions, agent modes, Git-aware actions, and tool approvals.

Desktop

Download Truss Desktop for Windows or Linux. Open a workspace, choose Settings, configure a model endpoint, and select a model. The desktop app includes files, editor and diff views, a terminal, Git controls, persistent chat, MCP settings, and approval controls.

Quick start

1. Start a compatible model server. Ollama is the simplest option; LM Studio,
   llama.cpp, and other OpenAI-compatible servers also work.
2. Install the CLI or terminal UI, or install the VS Code extension or desktop
   app.
3. Let Truss discover your local server, or configure a profile explicitly.

Ollama is detected at http://127.0.0.1:11434. Compatible servers are also
probed at common LM Studio and llama.cpp addresses. Check what Truss finds:

truss-cli models

Configuration

The CLI and terminal UI share JSON configuration files. Run these commands to
see the active user and workspace locations, then create a safe workspace
template:

truss-cli config path
truss-cli config init

Example .truss-harness/config.json:

{
  "defaultProfile": "ollama-coder",
  "profiles": {
    "ollama-coder": {
      "provider": "ollama",
      "baseUrl": "http://127.0.0.1:11434",
      "model": "qwen3:8b",
      "mode": "edit",
      "permission": "ask",
      "internetAccess": false
    },
    "lm-studio-fast": {
      "provider": "openai-compatible",
      "baseUrl": "http://127.0.0.1:1234/v1",
      "model": "local-model-id",
      "mode": "plan",
      "permission": "auto-read"
    }
  }
}

Use a profile for a task:

truss-cli chat --profile ollama-coder "Explain the runtime architecture"
truss-cli chat --profile lm-studio-fast --mode edit "Update the tests"

You can also configure a one-off session with environment variables:

The Desktop app and VS Code extension keep endpoint, model, mode, and
permission choices in their own settings. API keys should stay in the client’s
credential flow or environment variables—never commit them to a workspace
configuration file.

Modes and approvals

Choose the mode that fits the task:

Tool permissions are independent from mode. Use Ask every time to approve
each action, Auto-allow read-only to approve safe inspection automatically,
or Auto-allow all tools only in a workspace you trust.

MCP servers

Truss can load local Model Context Protocol servers over stdio. Add them to
your shared CLI/TUI configuration, or paste the same object into the Desktop or
VS Code MCP settings:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "."],
      "cwd": ".",
      "enabled": true,
      "readOnly": true
    }
  }
}

Chat mode does not load MCP tools. Plan mode loads only servers marked
readOnly; Edit mode loads every enabled server. Workspace MCP definitions are
ignored by default because they can start local processes; explicitly trust
them with "allowWorkspaceMcpServers": true in your user configuration when
appropriate.

Learn more

• Documentation
• Getting started
• Client guides
• Download Truss Desktop
• GitHub issues

Truss is actively developed. Contributions and feedback are welcome.