Agent Vision — diagrams for the user

Agent Vision is a hosted diagram app. You write a small JSON document (the spec — a DSL), the app renders it (2D now, isometric later), stores every change as an immutable revision, and returns a public share URL plus a rendered PNG so you can show the user immediately.

This installed skill is a thin connection recipe. The full DSL, the color / idiom grammar, worked examples, ops-patch reference, icon catalog, and warning list live in the hosted, canonical skill doc — fetch it once at the start of a diagramming session and follow it:

<app-url>/skill.md   (ask the user for their Agent Vision app URL)

(Replace the host with whatever app URL the user gave you; the skill doc lives at <app-url>/skill.md.)

Connecting (given only the user's app <url>)

Step 0 — check the environment first: if AGENT_NATIVE_VISION_URL and AGENT_NATIVE_VISION_TOKEN env vars are set (provisioned containers), use them and skip all auth below — bearer header on every call, done.

Primary — device flow, you drive it. No pre-minted token, no password:

1. POST <url>/_agent-native/mcp/connect/device/start (empty JSON body, no auth) → { device_code, user_code, verification_uri_complete, interval, expires_in }.
2. Tell the user in chat: "Open <verification_uri_complete> and click Authorize this device (code <user_code>)." Any device where they're already logged in works.
3. POST <url>/_agent-native/mcp/connect/device/poll with { "device_code": "<device_code>" } every interval seconds until it returns approved → { token, mcpUrl, ... } (pending = keep going; expired/consumed = restart at step 1).
4. Use the token exactly as below.

Fallback — user-minted token: the user logs in, opens <url>/_agent-native/mcp/connect, clicks Create connection token, and pastes { url, token } to you. Tokens are per-user, scoped, expiry-bound, and revocable from that same page.

MCP clients (Claude Code, etc.):

claude mcp add --transport http agent-vision \
  <url>/_agent-native/mcp \
  --header "Authorization: Bearer <token>"

OAuth-capable clients (Claude web/Desktop, Cursor, ChatGPT): add just <url>/_agent-native/mcp (no header) and approve the connector in the client's sign-in popup — no token needed.

Raw-HTTP agents (no MCP client): every tool is also a plain HTTP endpoint at <url>/_agent-native/actions/<name>, and the same bearer token works there. Send Authorization: Bearer <token> on every request. Reads are GET with query params; writes are POST with a JSON body. Without the header these routes return 401 {"error":"Unauthorized"}.

First contact (5-line recipe)

1. list-projects                         → discover/confirm the workspace
2. create-diagram { title, spec }        → returns { url, svgUrl, pngUrl, warnings }
3. fetch pngUrl                          → LOOK at the render; read warnings
4. update-diagram { diagram, ops:[...] } → fix warnings (repeat 3–4 as needed)
5. share url                             → send the /d/<slug> link to the user

• Never share a diagram you haven't rendered and looked at, or one with outstanding warnings — warnings are a todo list, not decoration.
• The share url (<app-url>/d/<slug>) is public and needs no token; that is the link you hand the user. .svg / .png variants are the raw images.
• For anything beyond this loop — the spec shape, node/edge/group fields, ops patches, icon names, idioms (callouts, zones, trunk edges, code text) — read <app-url>/skill.md, which is the single source of truth and stays current with the deployed app.