Agents

Steering an agent means writing: the spec it implements, the plan it follows, the AGENTS.md it reads first. In Sundial those documents and the agents share one workspace, so the steering document is never out of the agent's sight, and the agent's work is never out of yours.

Sundial Agent, built in

Sundial Agent is the embedded agent, one per chat. It reads and edits workspace documents directly, and its edits land in your chosen mode: direct edits, suggestions, or read-only answers. For heavier work it has a real computer: a per-agent sandbox where it installs packages, runs scripts, compiles LaTeX, and works with git, booted on demand and idled down when unused. It can also comment, pin checkpoints, search version history, and schedule its own follow-ups.

Connect your own agent

Claude Code, Codex, Cursor, and any MCP client join a workspace through Sundial's hosted MCP server. A connected agent appears in the presence bar, shows a live cursor while it works, and every one of its edits is tracked under its name.

claude mcp add --transport http sundial "https://www.sundial.md/mcp?workspace=<id>"

codex mcp add sundial --url "https://www.sundial.md/mcp?workspace=<id>" && codex mcp login sundial

Connected-agent writes default to suggestions, so their work waits for your review. Each agent has a “suggest only” toggle if you want to hold it there.

Harnesses

Each chat picks the harness Sundial Agent runs on: the default runs any model through the Vercel AI SDK, or switch a chat to the Claude Agent SDK or the OpenAI Agents SDK to get those harnesses' native behavior with their provider's models. The pick is per chat, in the model menu.

Models

Model selection is per chat too, from a catalog of frontier and open-source models across Anthropic, OpenAI, Google, xAI, and others. Swap models mid-session without losing the thread. Bring your own API key per provider in settings: chats on that provider's models then run on your key instead of Sundial credits.

Scheduled tasks

A scheduled task is a chat with a schedule: run once at a time, on an interval, or on a cron expression. Use it for the recurring writing work: the weekly status report drafted from the workspace, the changelog kept current. Results land as tracked changes like any other agent work.

iMessage

A workspace is also reachable from iMessage and RCS. Text Sundial Agent and it works in the same files, with the same attribution; open the workspace later to review what it did. The web editor is the primary surface, messaging is the one you have in your pocket.

Next steps