Connect once.
Absorb everything.
Downlink is the channel from your AI back to you — hosted pages you skim, maps you explore, audio for the walk, and questions that find your phone. This page covers everything: connecting, what your agents can make, and how the loop back to them works.
Quickstart
Claude Code — one command:
Any other MCP client (Codex, Cursor, claude.ai connectors, your own agent) — add a streamable-HTTP MCP server pointed at:
The first time your agent calls a Downlink tool, you'll be asked to sign in (Google or email) — once per client. That's the whole setup. No SDK, no config files, nothing to install in your project.
Bearer token instead.Publish your first page
You don't call Downlink — your agent does. Just ask for the result you want:
The agent publishes and hands you a URL like
downlink.fyi/s/… — open it anywhere, it's built for your phone.
Better: agents that have Downlink connected will start offering it at the right
moments (long task finishing, you stepping away) without being asked.
Every artifact is private to you by default — the URL alone opens for no one else. More under Sharing & privacy.
What your agents can make
Four formats, each matched to how you actually absorb:
| Format | What it is | Reach for it when |
|---|---|---|
| Page | A hosted, interactive HTML page — results, comparisons, dashboards, write-ups, with real product screenshots if the agent has them. | You'd rather skim something visual than read a wall of terminal text. |
| Mind map | An explorable, zoomable tree of a topic, codebase, or plan. Branches expand as you click. | Structure matters more than prose — architectures, decisions, research. |
| Narration | An audio briefing written for the ear, with transcript. Plays in the browser. | You're walking, driving, or done with screens. |
| Question queue | A phone-friendly list of questions — multiple choice, free text, approvals, FYIs — you answer whenever. | The agent has questions but shouldn't stop working to ask them. |
Artifacts update in place: when the agent revises one, the same URL shows the new version, live — no new links to hunt for.
Talking back
Every Downlink artifact is two-way. What comes down as a page goes back up as answers:
- Interactive pages — agents wire real buttons and inputs into their pages (pick an option, approve a plan, fill a form). Your clicks flow straight back.
- Question queues — answer, skip, or comment on any question from your phone. Answered questions archive themselves; the queue only ever shows what still needs you. Your agent keeps working the whole time and picks up answers when it polls.
- Annotate — comment on any element of any page or map, even when the agent added no inputs at all. See below.
Agents read all of it with one tool (get_responses) — no webhooks, no
infrastructure, works from any machine the agent runs on.
Click anything, say anything
Every page and mind map has an ✏️ Annotate button (bottom-right). Click it, then click any element — a paragraph, a chart, a mind-map node — and write a comment. Enter sends it.
Your agent receives the comment anchored to what you clicked: on a mind map, "Root › Auth › Token refresh"; on a page, the exact element and its text. It's the universal reply channel — the agent doesn't have to anticipate what you'll want to react to.
Your dashboard
app.usedownlink.com is the one place with everything:
- Artifacts by project — everything your agents have published, grouped by the work it belongs to, with version history.
- Questions strip — every queue with something waiting for you, across all projects, front and center.
- Settings — your API key and preferences (like the default narration voice).
Tool reference
What your agent sees once connected. You never need to call these yourself — this is for the curious and for prompting precisely.
| Tool | What it does |
|---|---|
| downlink_publish_page | Publish a self-contained interactive HTML page. Pass id to revise in place. |
| downlink_publish_mindmap | Publish a markdown outline as an explorable mind map. |
| downlink_narrate | Turn an ear-authored script into a hosted audio briefing (five voices). |
| downlink_ask | Add non-blocking questions to the project's queue; withdraw moot ones. |
| downlink_get_responses | Pull back everything the human did: opens, answers, clicks, annotations. |
| downlink_mint_links | Create labeled share links — access grant + attribution per recipient. |
| downlink_stage | Upload a large page out-of-band (zero context cost), then publish by handle. |
| downlink_check_page | Lint page HTML against the sandbox before or after publishing. |
| downlink_get_artifact · list · edit · delete | Lifecycle: fetch source, list by project, find-replace edit, delete. |
Pages run in a sandboxed frame with a strict security policy: no network calls out, no data leaving the page except through the response channel you see.
FAQ
Is it free?
Yes — Downlink is in early access, free while we shape it with the people using it.
Which agents work with it?
Anything that speaks MCP: Claude Code, claude.ai, Codex, Cursor, and custom agents. One account, all your agents — everything lands in the same dashboard.
Do my agents need special prompting?
No. The tools teach themselves — agents discover when to publish, ask, and poll from the tool descriptions. Saying "put it on Downlink" works on day one; by day two they offer it themselves.
What happens to a question I don't answer?
It waits. Nothing expires out from under you; agents treat silence as "not yet", not "no". Skipping is also an answer — it tells the agent you saw it and passed.
Where does my data live?
Artifacts are stored encrypted at rest on Cloudflare's edge, served over HTTPS, private by default, and deletable by you at any time.