← All guides

Connect an assistant to CraterView MCP server

CraterView as an MCP server — add one address to Claude or any MCP client, paste a code from your dashboard, and your assistant can enhance and restore photos.

Last updated 16 September 2026

If you would rather say “sharpen this and make it twice the size” than write a request, an assistant can do the work for you. CraterView has an MCP server, which is the open protocol assistants use to reach tools, and connecting one takes a signed-in account, one address and a code. No API key changes hands.

https://mcp.craterview.ai/mcp

1 — Sign in first

An assistant connects to an account, and a guest workspace is not one you can come back to: it lives in a single browser’s cookie, so an assistant linked to it could never be found or disconnected from anywhere else. The dashboard’s Connect an assistant panel says as much and offers sign-in instead of a code until you have.

2 — Add the server to your client

Claude Code, from the terminal:

claude mcp add --transport http craterview https://mcp.craterview.ai/mcp

Or install the plugin, which adds the server and a skill that knows which tool to reach for:

/plugin marketplace add craterviewai/craterview-python
/plugin install craterview@craterviewai

Claude on the web or desktop: Settings → Connectors → add a custom connector with the address above.

Any other MCP client: add it as a remote server over HTTP at that address. The connection is authorized the same way everywhere.

3 — Paste the code

The first time the client uses the server it opens a CraterView page asking for a code. In your dashboard, open Connect an assistant and press Get a code; paste what it gives you. A code works once and for ten minutes, and it is all the page ever asks for — a key is never shown, typed or copied on this path.

From then on every job the assistant runs is on your account: billed to it, and in your history like any other.

4 — Ask for the work

The server offers two tools, and the assistant chooses between them:

• enhance_image takes one image — a public URL, or the image itself if it is already in the conversation — and a model, and runs it. Most images come back finished inside the call, with a link to the result.
• get_job collects a job that was still running when enhance_image returned, using the job id it was given. The assistant should wait a few seconds between calls rather than ask again; asking enhance_image again would run and charge the work twice.

So the prompts are the ordinary ones: “upscale this to four times the size”, “this scan is soft, sharpen it”, “this print is torn and faded — repair it”. The assistant picks the enhancing model for the first two and the restoring model for the last; if the request could mean either, a good assistant asks, because enlarging a damaged print sharpens the damage.

A result is a link, not the bytes, and the link expires. Use it, or fetch the image, when it arrives; get_job mints a fresh one if you come back later.

What it costs, and where it shows

A job an assistant runs is charged exactly as one you submit yourself: the price is per model and published by the catalogue, and a failed job is not charged. An account with no credit still runs its work on the community queue, which is served after paid work — a longer wait rather than a refusal. Everything the assistant did is in your dashboard’s history, so a surprise is one click from an explanation.

Disconnecting

Remove the server from your client and the connection is gone. Rotating your API key from the dashboard also disconnects every assistant linked to the account at once, because the connection stands on that key — which is what to do if you no longer trust a client that had one. An assistant whose connection was revoked is told so and asked to reconnect; a fresh code is all it needs.

Three things that catch people out

An image pasted into the conversation is expensive to send. The assistant has to pass it inline, and a one-megabyte image is hundreds of thousands of tokens of context. Give it a URL whenever the image has one.

A local file has no URL. If your client can read your files but not serve them, inline is the only route — send the smallest copy that will do.

“Still running” is not “try again”. A large image takes tens of seconds. The assistant should collect it with get_job; if it calls enhance_image a second time, that is a second job and a second charge.

Next

If you would rather have the assistant write the integration than run it — a script, a service, a batch — the agent skills teach it the API, and your first API call is the same three requests by hand.