<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[Cirrela]]></title><description><![CDATA[Cirrela is an independent French product by Lorenzo Corpita. We make personalised video wishes — Santa, Tooth Fairy, Three Kings say a child's first name — and ]]></description><link>https://cirrela.hashnode.dev</link><image><url>https://cdn.hashnode.com/uploads/logos/6af9d125206e155cd62b7a44/e73412e3-3b3a-4140-8628-10fe81a62441.png</url><title>Cirrela</title><link>https://cirrela.hashnode.dev</link></image><generator>RSS for Node</generator><lastBuildDate>Wed, 18 Nov 2026 14:14:15 GMT</lastBuildDate><atom:link href="https://cirrela.hashnode.dev/rss.xml" rel="self" type="application/rss+xml"/><language><![CDATA[en]]></language><ttl>60</ttl><item><title><![CDATA[How we built an MCP server that turns a sentence into a video]]></title><description><![CDATA[At Cirrela we sell AI-generated videos — most famously a 10-second clip of Santa saying your kid's name. We also expose the whole studio to AI agents…]]></description><link>https://cirrela.hashnode.dev/how-we-built-an-mcp-server-that-turns-a-sentence-into-a-video</link><guid isPermaLink="true">https://cirrela.hashnode.dev/how-we-built-an-mcp-server-that-turns-a-sentence-into-a-video</guid><category><![CDATA[mcp]]></category><category><![CDATA[AI]]></category><category><![CDATA[video]]></category><category><![CDATA[SaaS]]></category><dc:creator><![CDATA[hello]]></dc:creator><pubDate>Sun, 15 Nov 2026 15:41:09 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/covers/6af9d125206e155cd62b7a44/78e597c7-6789-4b76-8dfe-39d02870c105.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>At Cirrela we sell AI-generated videos — most famously a 10-second clip of Santa saying your kid's name. We also expose the whole studio to AI agents through an MCP server. Here's how it works under the hood.</p>
<p><strong>Transport: Streamable HTTP</strong></p>
<p>Our MCP endpoint is <code>https://cirrela.com/api/mcp</code> — a single Streamable HTTP endpoint (the modern replacement for the old HTTP+SSE transport). Clients POST JSON-RPC messages and get either a JSON response or an SSE stream for server-initiated messages. One URL, works with Claude, Cursor, and any MCP client.</p>
<p><strong>Auth: OAuth 2.1</strong></p>
<p>The server sits behind OAuth. When a client connects, it goes through a standard authorization-code flow with PKCE — the user logs in with their Cirrela account and grants scopes. Tokens are bearer tokens; the MCP layer maps them to the same account and credit balance the web app uses. No API keys to paste.</p>
<p><strong>Tools</strong></p>
<p>We expose a small, opinionated toolset rather than mirroring our whole API:</p>
<ul>
<li><code>create_video_wish</code> — the flagship. Given a child's name, a character (Santa, elf, Tooth Fairy…), and a language, it queues a render and returns a job ID.</li>
<li><code>list_models</code> / <code>get_model</code> — the studio side: around 9 video models (image-to-video, text-to-video) with per-model pricing in credits.</li>
<li><code>get_job</code> / <code>list_jobs</code> — poll render status.</li>
<li><code>get_credits</code> — balance and plan info.</li>
</ul>
<p>Each tool returns structured JSON, and the descriptions tell the agent the price <em>before</em> it spends — we deliberately made agents quote costs up front.</p>
<p><strong>Credits, not per-call billing</strong></p>
<p>Renders consume credits from a subscription (Starter / Pro / Studio). Credits never expire while subscribed. Failed renders are automatically refunded — the job status flips to <code>failed</code> and the credits come back, so an agent can retry safely.</p>
<p><strong>Why MCP</strong></p>
<p>The interesting part isn't the protocol — it's that agents become distribution. When Claude can call <code>create_video_wish</code> directly, "make a video of Santa saying my daughter's name" turns into a two-tool-call task. The MCP server is also how we dogfood: our own agent flows hit the same endpoint customers do.</p>
<p>The whole thing is a few hundred lines of handler code on top of our existing backend — the hard part was making tool schemas descriptive enough that an LLM picks the right model and warns the user about cost before rendering.</p>
<p>That's it: one sentence in, one video out.</p>
]]></content:encoded></item></channel></rss>