Remote hosted MCP servers are the 2026 default

The shift from local processes to a hosted endpoint

For the first couple of years of MCP, most servers ran as a local process. You installed a package, kept it running, and your agent talked to it over stdio on the same machine. That was fine for reading files or hitting a database. It stopped making sense the moment the work needed real compute.

In 2026 the default flipped. Clients now expect you to add a server by pasting a URL and signing in. The Streamable HTTP transport is stable, remote connectors are a first-class option in every major client, and a local install has become the exception rather than the rule.

Media generation is the clearest reason why. You are not shipping a small tool. You are asking for frames from an image model and a video model, and that is not something you want tied to your own laptop.

Media models do not belong on your laptop

Image and video models are large and they change often. Running them locally means downloading weights, keeping a GPU busy, and re-downloading every time a new version lands. Most developer machines cannot do it at all, and the ones that can pay for it in heat and time.

A remote server moves that whole problem off your desk. Figgshield runs Nano Banana 2 for images and Kling 3 for video, with a faster Kling 3 Turbo tier when you want speed over polish. You send a prompt, the server does the heavy part, and you get a file back.

Because the models live server-side, you are always on the current version. There is nothing to update and nothing to reinstall when the model improves.

One URL, every agent

A hosted endpoint means the same connector works everywhere. You add https://api.figgshield.ai/mcp once, sign in, and the tools show up in whatever agent you are using that day.

This matters more than it sounds. A designer might live in Claude while an engineer stays in Cursor, and both hit the same server, with the same credits and the same output waiting for them.

• Claude, both desktop and Claude Code
• Cursor
• Windsurf
• ChatGPT in Developer Mode (paid plans)
• any other MCP-capable agent

What Figgshield runs behind that endpoint

The connector exposes a small set of tools: generate an image, generate a video, and check what you have already made. An image from Nano Banana 2 costs 8 credits at 1K or 2K and 16 credits at 4K. A five-second video clip is 35 credits, whether you use Kling 3 or the Turbo tier.

One credit balance covers both image and video, so you are not juggling separate wallets. Pricing is at cost. We do not mark up the models, and the only money we make is on credits people buy and never spend. You can see the current numbers on the pricing page or check your credit balance at any time.

If you already generated an image you like, you can start a video from it by passing the image id. The server treats your existing output as the first frame, so you are not re-describing a scene you already have.

Files, references, and what disappears after seven days

Every file the server makes sits on a private link for seven days, then it is deleted. That is long enough to download it, drop it into a project, or hand it to a teammate, and short enough that nothing lingers on a server you forgot about.

A reference image you send in works differently. It is used in memory to guide the generation and is never stored. When the request finishes, it is gone. Nothing you upload becomes part of a dataset or sticks around on disk.

Moving over: a short checklist

If you are still running a local media tool, switching takes a few minutes.

New to it entirely? There is a 7-day Pro trial that includes 150 credits for $5. There is no free tier, but if you cancel before it renews you pay nothing more.

1. Open your agent's connector or MCP settings.
2. Add https://api.figgshield.ai/mcp as a remote server.
3. Sign in when the client opens the browser flow.
4. Send a test prompt and confirm the tools appear.
5. Remove the old local server so your agent stops reaching for it.