Run agents from the Terminal

1. Install and check the runtime

The published Terminal package requires Node.js 22.13.0 or newer. Install the package, then run setup in an interactive terminal to choose language, runtime preference, and default permissions.

npm install -g agentlas
agentlas version
agentlas setup
agentlas doctor
agentlas list

Sign in to the model CLI or configure the provider you intend to use. Agentlas account login and model-provider login are separate. doctor shows the runtime state on this machine; use an actual small task to check execution.

2. Connect the current folder to a project

Change to the folder where you want the work to happen. Choose an exact installed agent from agentlas list. project use connects that folder with one controller; project team saves an ordered team. In the current Terminal, the first member is the controller and must be locally installed.

agentlas project status
agentlas project use "<installed-agent-slug>"
agentlas project status

# To choose several installed agents:
agentlas project team "<controller-slug>" "<specialist-slug>"

These commands save the project's team; they do not merely choose a model for one call. Inspect the printed project and controller before running. If the folder or team is ambiguous, resolve the reported problem instead of assuming another agent will take over.

Optional: agentlas project init prepares local Agentlas project context for Context Map features. It can create .agentlas/ files and update .gitignore. It is a separate action from connecting the project's team.

3. Run a small task

agentlas run --permission read "Summarize the README in this folder and identify three points to check."

# Directly call an exact installed agent:
agentlas run "<installed-agent-slug>" --permission read "Summarize the supplied document."

Without an explicit agent, run uses the current folder's saved project controller. Supplying an exact installed agent selects a direct call. Use --runtime to choose an available runtime and --model for an exact provider model when supported. Check agentlas help run for your installed version; --print returns the final answer without the streaming interface.

Check the actual response and any output file. Read permission is a useful first test; choose write or full only when the intended task needs that access. A model selection or process exit is not a substitute for the requested result.

4. Run a saved graph

agentlas graph list
agentlas graph show "<name>"
agentlas graph run "<name>"

# When the graph requires a starting value:
agentlas graph run "<name>" --input "<your value>"

Review the steps and outward effects first. The current Terminal executes graph run directly through its execution core and returns its result or error. It does not simply leave a request for Desktop's next launch. Recurring schedules are managed separately from this one-time direct run.

5. Shared state and recovery

On the same computer, Terminal and Desktop use shared local project, agent, and runtime data. Their execution processes are separate. Check doctor if a native database driver, runtime, or compatible execution core is missing; keep the reported error and verify the repaired component before retrying.