Agents
Your agents are collaborators in the document: their edits are attributed, land as suggestions, and wait for your review.
Claude Code and Codex, on your machine
If Claude Code or Codex is installed, it is the default engine for every chat. The turn runs through the tool itself, with your own login and plan, and nothing goes through Sundial: no API key, no credits, no account. The composer shows which engine a chat runs on, and existing terminal sessions on your Mac can be imported as chats and resumed in the editor.
Pick the engine per chat in the model menu. A chat keeps its engine once it has messages, so to switch engines, start a new chat.
Sundial Agent, in the cloud
Sundial Agent is the hosted agent. It needs a Sundial account and runs on Sundial's servers, with a sandbox of its own for shell commands, so it never runs anything on your machine. It can use any model in the catalog (Anthropic, OpenAI, Google, xAI, and open models), switched mid-chat, and it is the engine behind shared documents, scheduled tasks, and the web app.
Sundial Agent runs on Sundial credits by default. You can add your own Anthropic, OpenAI, Google, or xAI key under Settings → API keys so chats on those providers run on your key, or connect your Claude or ChatGPT subscription there so cloud chats run on your plan.
Ways to ask
Chat (Cmd+Shift+J), with Cmd+J to add a selection. A comment starting with @Agent, answered in the thread with a suggestion on that selection. A chat set to watch a document's comments. Inline, with /ai or Cmd+G. All of them produce the same attributed, reviewable edits.
What an agent can do
Agents read, write, and search the project's files, reply to comments, add checkpoints, and look through history. In Suggest mode, the default, every write is a tracked change. In View mode the write tools are removed, not just discouraged, so the agent can only read and answer.
AGENTS.md at its root, that file is what gets edited.Connect an agent from outside
A workspace also accepts agents running elsewhere, and the way in is Open with…: the last icon in the sidebar's strip, also in the chat pane and the ⌘T launcher. It hands the agent a short prompt that points at this workspace's /startguide, and the agent joins over plain HTTP: it appears in the presence bar with a live cursor, and its edits are tracked under its name, as suggestions by default.
Web copies that prompt for any AI, or opens ChatGPT or Claude.ai with it already filled in. Desktop hands the same workspace to Claude Code (one click) or Codex (copy the prompt) running on your machine. Agents cannot reach a localhost URL, so both need the workspace on a public https address.
For a direct MCP connection from a terminal, Settings → Advanced → MCP connections generates the command with your workspace id filled in.
claude mcp add --transport http sundial "https://www.sundial.md/mcp?workspace=<id>" && claude mcp login sundial
codex mcp add sundial --url "https://www.sundial.md/mcp?workspace=<id>" && codex mcp login sundialSee the web app page for scheduled tasks and skills, which run on the hosted agent.