A remote MCP server URL is the single address that lets Claude reach tools like Figma, and this article shows real examples. You will see the exact Figma URL, the steps to add it in Claude.ai and Claude Code, and what is realistic for LinkedIn and Instagram, where the servers come from third parties.
A remote MCP server URL is one line of text, and it decides whether Claude can reach your tools or stays stuck inside a chat window. Paste the right address into the right field and Claude can read a Figma file, pull an Instagram insight, or draft a LinkedIn post from real data. Paste the wrong one and you stare at a "connecting" spinner that never resolves.
This article puts working examples side by side. You get the real Figma URL, the exact steps to add any URL to Claude.ai and Claude Code, and an honest look at LinkedIn and Instagram, where the servers come from third parties instead of the platforms themselves. A table of the mistakes that break connections follows, then a section on connecting PicassoIA's own MCP connector so Claude can generate images and videos from a conversation.
💡 Quick answer: a remote MCP server URL is an HTTPS address, usually ending in
/mcp, that speaks the Model Context Protocol and asks you to sign in with OAuth. Figma's ishttps://mcp.figma.com/mcp. Claude.ai takes it under Settings > Connectors > Add custom connector.
The Model Context Protocol, or MCP, is the open standard that lets an AI assistant call outside tools. An MCP server exposes those tools: read a design file, list recent posts, generate an image. An MCP client, such as Claude.ai or Claude Code, connects to the server and decides when each tool is worth calling.
A remote server lives on the internet and answers at a web address. That address is the whole setup. You do not install anything, run a process, or edit a config file on your own machine. When you paste the URL, the client connects, asks the server which tools it offers, and sends you through a sign-in if the server requires one. From then on, Claude can call those tools mid-conversation.
Take the Figma address, https://mcp.figma.com/mcp, and split it up:
/sse, which marks the earlier streaming transport.The path is the part people get wrong most often. Copy the full URL from the vendor's own docs, including the path, instead of retyping the host from memory.
Sign-in comes in three styles. Most servers use OAuth, where you approve access in a browser window. Some let Claude connect first and ask you to sign in only when a tool needs it. A few need no sign-in at all, or accept an access token sent as a request header. Figma uses OAuth, so the first connection always opens a browser.
| Remote server | Local server | |
|---|---|---|
| What you enter | A URL | A command that launches a process |
| Where it runs | The vendor's or a host's infrastructure | Your own computer |
| Sign-in | OAuth in the browser | Often a token in an environment variable |
| Claude.ai custom connectors | Supported | Not an option, the field expects a URL |
| Updates | Handled by the server owner | You update the package yourself |
Local servers are fine for private files and quick experiments. For anything you share with a team, a remote URL is easier to roll out, because nobody has to install a thing.
The steps differ slightly by plan and by app, but the idea is identical: give Claude the URL, sign in once, then switch the connector on in the chat where you want it.
On Free, Pro and Max plans:
The Free plan allows a single custom connector, so pick the one you will use most.
Claude Code adds servers from the terminal. For Figma, one command does it:
claude mcp add --transport http figma https://mcp.figma.com/mcp
Then type /mcp inside Claude Code and follow the browser prompt to authorize. Figma also offers a plugin route, claude plugin install figma@claude-plugins-official, which it lists as the preferred method. The manual command above is the one to remember, because it works for any remote server once you swap in the right URL.
Two people are involved. An owner adds the connector under Admin settings > Connectors with the same URL. Each member then opens Settings > Connectors, finds the entry marked Custom, and clicks Connect to sign in with their own account. That split matters: the owner shares the address, but every member keeps their own permissions.
💡 Treat custom connectors like installing software. A custom connector can reach unverified services, and Claude can take actions inside them. Connect only to organizations you trust, switch off connectors you are not using, and stay alert for prompt injection, where text inside a file or post tries to steer Claude.
Figma is the cleanest example of a remote MCP server done properly: a stable public URL, OAuth sign-in, and a clear list of supported clients.
The address is https://mcp.figma.com/mcp. The first time you connect, your client opens a browser window, you approve access, and the client reuses those credentials later. Figma states that only clients listed in its MCP catalog can connect, and Claude Code, VS Code and Cursor are named among them. If a client is missing from that catalog, the URL alone will not get you in.
Not every client has a Connectors screen. In VS Code you open the user MCP configuration and paste:
{
"servers": {
"figma": {
"url": "https://mcp.figma.com/mcp",
"type": "http"
}
}
}
Codex has a command: codex mcp add figma --url https://mcp.figma.com/mcp. Notice that the URL never changes. Only the wrapper around it does, and that is the real lesson of remote MCP: keep one address and reuse it in every client.
With the connection live, ask for concrete work tied to a real frame:
Paste a frame link in your message so Claude knows exactly which design you mean. Vague requests like "make the homepage" waste tool calls and return generic results.
Here the honest answer changes. Figma publishes its own server. LinkedIn and Instagram, as far as I could find, do not publish official remote MCP servers, so every URL you see for them belongs to a third party. Check each vendor's docs for changes, because this area moves quickly.
The most popular LinkedIn MCP project runs on your own machine and drives a Playwright-controlled Chromium browser through your logged-in session. That makes it a local server, not a remote URL, and LinkedIn's terms of service prohibit scraping, so it carries account risk. Remote options exist, but they come from other companies.
| Route | How it connects | Risk level |
|---|---|---|
| API-based server | Official Community Management API, mainly for Company Pages | Lowest |
| Aggregator bridge | A hosted URL from a service such as Zapier or Composio | Medium, depends on the provider |
| Browser automation | Local server using your logged-in session | Highest, can trigger account restrictions |
The variable that matters is whether the server acts through LinkedIn's API with your authorization, or drives your session with a scraped cookie. Prefer the first. A remote example looks like this, with a placeholder you replace with your provider's real address:
https://your-linkedin-bridge.example.com/mcp
Good first prompts are read-only: "Summarize the last ten posts on our company page and group them by topic." Hold off on write actions until you have checked what the server can post on your behalf.
Instagram servers sit on top of Meta's Instagram Graph API, which only works with Business or Creator accounts. A personal account cannot use it, and that single fact explains most failed setups. Some remote servers are read-only, exposing profile data, media, insights and comments. Others, such as the community server from luminarylane, can also publish photos, carousels and reels.
The pattern matches Figma, only the URL is a placeholder:
claude mcp add --transport http instagram https://your-instagram-server.example.com/mcp
Sign in with the Meta account tied to your professional profile and grant the permissions the server requests. Then try: "List my five most recent posts with likes and comments, and tell me which format performed best." Always ask a read-only question first, since a publishing tool can post to a live audience.
| Symptom | Likely cause | Fix |
|---|---|---|
| Spinner never ends | Wrong path, such as /sse on a server that expects /mcp | Copy the exact URL from the vendor docs |
| "Cannot connect" | http:// instead of https:// | Use the secure address |
| Works in Claude Code, fails in Claude.ai | Address is localhost or on a private network | Host it publicly, or use Anthropic's MCP tunnels |
| Sign-in loops | Pop-up blocked, or wrong account | Allow pop-ups, then retry with the right account |
| Tools missing in chat | Connector not switched on | Press +, open Connectors, toggle it |
| Instagram returns nothing | Personal account | Switch to a Business or Creator account |
When a URL fails, check it outside Claude. The official MCP Inspector connects to a server directly and lists its tools, so you can see whether the server or the client is the problem. If Inspector connects and Claude does not, look at the connector toggle and your plan limits. If Inspector fails too, the URL or the server is at fault.
Everything above is about other companies' servers. PicassoIA also offers an MCP connector, which helps when you want Claude to produce visuals for the posts it drafts.
The connector gives Claude tools to generate and edit media on PicassoIA's own GPUs: generate_image, edit_image, generate_video_picassoia and generate_video_seedance, plus get_generation, cancel_generation, list_generations, list_models and get_account. Four models sit behind it:
Generation is asynchronous. A generate call returns a prediction ID right away, then Claude polls get_generation until the job succeeds or fails. Up to five predictions can run at once per account, shared across every connection.
The connector URL is not printed on the public site. Sign in and open the MCP connections page at picassoia.com/en/mcp/accounts to copy it. Do not paste the REST base address, https://api.picassoia.com/v1, into the connector field unless that page tells you to: the REST API is a separate route for direct HTTP calls with a Bearer token. Check the pricing page to see which plan includes MCP connections before you set it up.
Once it is added, a first request can be as short as: "Generate a 16:9 photo for this post and show me the result." Claude calls generate_image, polls until the job finishes, and returns the image link.
When a connector fails, you often need a second pair of eyes on the error text. Claude Sonnet 5 on PicassoIA reads code, accepts screenshots, and works through multi-step tasks, which suits connection debugging well.
low, fine for quick questions. Choose high for tangled setups, up to max for the stubborn ones.For a second opinion on a hard case, Claude Opus 4.7 handles heavier reasoning, while Gemini 3.5 Flash returns quick answers when you only need a short check.
💡 Ask for the matching command. Tell the model which client you use and which URL you have, then ask for the exact add command or JSON snippet for that client. Paste it, authorize, and test with a read-only prompt.
Pick one post idea from the Instagram example and turn it into visuals today. Open PicassoIA Image on Picasso IA, write a prompt with a clear subject, a light direction and a lens, and ask for a 16:9 frame. Refine the result in PicassoIA Image Editor Pro, then animate your favorite frame with PicassoIA Video to get a short clip for a reel.
If you prefer other looks, test Seedream 4.5 and Flux 2 Pro with the same prompt and compare the results side by side. Then connect the PicassoIA connector to Claude and let it do the whole loop from a single conversation. Experiment with three prompts, keep the best one, and build your own library of reliable prompts from there.
