Overview

Mercury is a command-line application for working with AI coding agents in a project folder. An agent reads, edits, runs and verifies code in the folder, several sessions run side by side, and the provider and model change without leaving the chat. Mercury runs as a full-screen terminal application, from an editor, headless in scripts, or on a schedule.

A session keeps its conversation, model, permissions, workspace and running work. Changing screens does not end the session. Opening another session leaves the previous session running.

Interface

The interface contains the following screens and controls:

• Boot face: The screen where an interactive boot with no explicit journey lands. Its ten rows are New Session, Continue Last Session, Boot Menu, MCPs & Skills, Agents, Doctor / Health Check, Saturn Scheduler, Logins, Session Concourse and Sessions · Projects.
• Chat: The view of the focused session. It shows the conversation and each tool call as it runs, as a compact card or its full output.
• Session Concourse: The board of the current project, with running sessions above parked chats. Each live row shows what its session is doing.
• The strip: The row of screens that exist. Shift+← and Shift+→ move between them. The chat joins the strip when a session is focused.

Continue Last Session appears once session history exists. Every row except Session Concourse opens in place as a layer of the Boot face, and Esc returns to the row. Session Concourse opens the board as a separate screen.

Installation and providers

Release archives carry their own Node runtime and ripgrep, so a release installation needs git only.

• Provider accounts: The sign-in catalogue covers OpenAI, Anthropic, OpenRouter, Google, Hugging Face, Moonshot, Z.AI and DeepSeek. /logins opens the catalogue, and /accounts manages the provider slots afterwards.
• Local servers and custom endpoints: Local model servers and a custom OpenAI-compatible endpoint need no sign-in. They become ready through discovery or configuration.
• Default model: An explicit --model or --effort on the command line wins. Without the flag, a new session starts on the choice saved by /model and /effort. With nothing saved, a fresh session starts on the provider of the most recent sign-in, on the newest model that sign-in can use, and a provider with no usable model falls through to the next most recent sign-in.
• Session controls: /model changes the session's model, saves the pick as the default and says which model was selected and why. /effort adjusts the session's effort and saves the pick the same way. /model default clears the saved choice.

Mercury asks for workspace trust after the first-run choices. Nothing requested by workspace configuration runs before trust is granted, and a grant covers the whole repository. Declining the trust dialog exits Mercury.

Start a session

To start a session, follow these steps:

1. Open a terminal in the project folder.
2. Run mercury.
3. On the first run, choose an appearance.
4. On the first run, sign in to a provider or choose sign in later.
5. Accept the trust dialog when it appears.
6. On the Boot face, select New Session.
7. Press Enter.

The blank chat opens on the model shown in the chip, and the session appears on the board.

Return to a session

To return to a session, follow these steps:

1. Enter /concourse in the chat.
2. Select the session's row.
3. Press Enter.

The selected session opens in the chat while every other session keeps running.

Related pages

Quickstart, Using Mercury, Sessions, Providers and sign-in, Models and effort, Permission modes, Install and update