Sundial for agents

What you are looking at

A Sundial workspace is a folder of files that humans and agents edit together. Every file is a live document. Every edit is attributed to whoever made it, human or agent, and an agent's edits arrive as suggestions the human accepts or rejects. You are a visible participant with a presence chip and a cursor, not a background script.

These docs are plain markdown. Fetch /docs.md for every page in one file, or /docs.md?page=for-agents for one page.

How you arrived, and what your credential is

You have	Do this
Claude Code or Codex launched by the desktop app	Nothing. You are that chat's agent. Write with your normal tools and Sundial tracks the edits.
A /w/<slug>?anon=<key> link, which is what Open with copies	Fetch /start. anon is your key: key= on every /g/ URL, or Cookie: sd_anon=<key> on the join call.
A pasted join prompt with the token and workspace id inline	Run its first command. It is the contract.
An Open with prompt: "Fetch /start and follow it, then join my workspace at <url>"	If that URL carries anon=, that is your key. If it carries nothing, you hold no credential yet. Say so and ask the human for the anon= link, a join prompt, or an MCP connection after they sign in.
A bare /w/<slug>	The same: no credential yet.
Nothing yet	/start creates a live, writable workspace in two calls, no account needed.

Six facts that unblock the first call

1. Call https://www.sundial.md exactly. The apex redirects and curl drops your headers on the way, so the wrong host is a 401. A localhost workspace is unreachable from outside.
2. Two headers on every request: Authorization: Bearer <token> and X-Agent-Id: ai:<your-name>, kept stable. Omit the id and you are ai:local-agent, the same name as every other agent that omitted it.
3. workspaceId is the UUID; a slug returns 401. POST /api/workspace/local-agent/join with {"projectId":"<slug>"} and your sd_anon cookie returns the UUID as workspaceId and a token good for 7 days.
4. Every write is a suggestion unless its body carries "editMode":"edit", and only when the human asked for direct edits. Join with {"editMode":"suggest"} unless they did; that locks your token. The human can also switch you to Suggest only from your chip, which blocks delete, rename, exec and uploads.
5. The token is a secret. Never paste it into chat or into a file. The link that carries anon= or token= is the same: never write it into a file, a commit, or a shared channel.
6. Fetch /start with whatever tool you have. A browsing tool gets the GET-only edition, where every operation is a URL under /g/ and your key rides in the query. A shell gets the full API. /start.md is always the markdown.

Connecting is one call, GET /api/workspace/local-agent/files?workspaceId=<uuid>: it registers your presence and returns the tree. Every endpoint is in /agent-docs; the workflow is in /start.

Feedback arrives on GET /events

Feedback arrives on GET /events: the human's comments on your work and their accept or reject on each suggestion, keyed by the suggestionId your write returned. Reply in the thread, revise, watch again. Never re-apply a rejected suggestion.

GET /file and GET /grep return the document as if every pending suggestion were accepted, yours and others'; a rejected suggestion's text disappears from later reads. Check GET /suggestions before anchoring an edit on text that might be someone's unreviewed proposal.

The rules

• Suggest by default. Every write is a suggestion unless its body carries "editMode":"edit" and the human asked for direct edits.
• Edit, do not overwrite. Anchor edits on exact existing text. A whole-file write on a document someone else is typing in loses their work.
• Say what you did. Report back in the chat or the thread you were asked in, briefly, and end with the question that moves the work forward.
• The token is a secret. Never paste it into chat or into a file.

MCP, once the human is signed in

MCP is for a signed-in human who wants a permanent connection to every workspace on their account; in a cloud workspace, Settings → Advanced → MCP connections prints the command. Start over HTTP.

Install the pointer

Install a pointer, never a snapshot. A downloaded copy of /start goes stale. The install stub in /start triggers on Sundial links and always follows the live version. It works for Claude Code at ~/.claude/skills/sundial/SKILL.md and for Codex at ~/.codex/skills/sundial/SKILL.md.

Where everything else is

• /start: the skill. Joining, creating, the local-folder live sync, and the template catalog.
• /agent-docs: the HTTP contract, every endpoint in full.
• /llms.txt: the signpost crawlers probe first.
• /templates.md: every template slug you can create a workspace from.
• /.well-known/agent.json: the discovery manifest.