Skip to content

The harness

Run agents on your machine with the Dexto CLI, using Dexto models or your own provider keys, with skills, MCP servers, and plugins.

The harness is the part of dexto that runs an agent on your computer: it talks to a model, calls tools, keeps chat history, and asks you before anything risky. It works signed out, with your own model provider keys, and gets more when you sign in.

Terminal window
dexto # chat in your terminal
dexto --prompt "review this diff" # start the chat with a prompt
dexto --continue # resume the most recent chat
dexto --resume <session-id> # resume a chat by id
dexto --agent ./my-agent.yml # use an agent from a file
dexto --mode web # the web UI

The first time you run it, dexto setup picks a provider and model. Run it again to change them. dexto connect connects a provider’s sign-in or key. Both have --no-interactive forms for scripts:

Terminal window
dexto setup --provider anthropic --model claude-sonnet-5-5 --no-interactive
dexto connect --provider <provider> --method <method> --no-interactive
Terminal window
dexto run "summarize this repository"
echo "fix lint errors" | dexto run

dexto run needs no terminal. It prints JSON when stdout is not a terminal, and the reply at one. Approvals nobody can give are refused and listed in the result, so a headless run never hangs on a question. See dexto run for the result and --stream-json events, and the reference for its options.

To let an agent run its tools without asking, use --auto-approve (or --permissions-mode auto-approve). Mandatory approvals still apply.

  • Your own key: dexto setup or dexto connect for a provider such as Anthropic, OpenAI or Google. You pay the provider.
  • Dexto models: signed in, models shown as dexto/<model> run on your sign-in and bill your organization’s credits in the current Cloud project. dexto models list shows them and dexto billing shows your balance. Spending credits asks for a permission the first time.

Pick a model with -m, --model for one run, or in the agent’s file.

Signed in, every local session gets a server called dexto with your connectors, Drive, Sites and Apps, Loops and memory. The same tools are available from the terminal with dexto workspace.

Terminal window
dexto mcp add github --command npx --arg=-y --arg @modelcontextprotocol/server-github
dexto mcp list
dexto mcp tools github
dexto mcp call github <tool> …

dexto mcp also configures an aggregation server. See MCP commands.

Terminal window
dexto agents list # installed and available agents
dexto agents install coding-agent
dexto plugin marketplace add owner/repo
dexto plugin install ./my-plugin
dexto init # make this folder a Dexto workspace
dexto skills push ./my-skill # add a skill folder to your library

See Local agents and plugins and Setup and workspace folders.

Terminal window
dexto chat list
dexto chat history <id>
dexto search "invoice"
dexto chat local # Claude Code and Codex chats on this machine
dexto chat import … # bring them into Dexto

An imported chat keeps which model answered each turn, and the answer shows it under More. When you continue the chat in Dexto, it uses the model of its last answer if your account can use it. Otherwise it uses your default model and says so. To continue Codex chats on your ChatGPT plan, run dexto connect codex first, or connect it later and import the chats again. See Continue imported chats on their model.