Remote MCP servers run at an HTTPS address, so Claude and ChatGPT connect with a pasted URL and no install. See five endpoints that answered without a login on 6 October 2026, the exact setup steps for both apps, a test command, and the security rules for free servers.
You do not need Node, Python, Docker or a config file to give Claude or ChatGPT new tools. A remote MCP server lives at an HTTPS address, and connecting it takes one pasted URL. The catch is that most "free MCP server" lists still point to local packages that need an install, a terminal and a JSON file that breaks when you miss a comma. This article sticks to servers you can reach right now. On 6 October 2026 I ran a Model Context Protocol handshake against each URL below from a plain terminal, and five of them answered without any login. One popular entry on older lists did not, and I will show you which one.
You will get the five working URLs, exact steps for Claude and ChatGPT, a test command you can point at any server, and the security rules worth following before you paste anything into a connector form.
The Model Context Protocol (MCP) is the open standard that lets an AI chat app call outside tools: search a docs site, read a repository, query a database. A server exposes the tools. The chat app is the client, and it decides when to call them.
A local server runs on your own machine over stdio. Your client launches it as a child process, so you need the runtime installed, the package downloaded and an entry in a config file. That works well for developers and badly for everyone else. It also means the server only exists on the computer where you set it up.
A remote server runs on someone else's infrastructure and talks over HTTPS. The modern transport is Streamable HTTP, usually served from a path that ends in /mcp. The older SSE transport, usually at /sse, still shows up on old lists and is being phased out. Prefer /mcp endpoints whenever a provider offers both.
One more distinction matters. Some remote servers are open, meaning anyone can call them. Others use OAuth, meaning the app sends you to a sign-in page the first time. This article sticks to open servers because they work with a single paste and no account setup. OAuth servers, such as the ones from GitHub or Notion, follow the same steps plus one consent screen.
| Local (stdio) | Remote (HTTPS) | |
|---|---|---|
| Install | Runtime, package and config file | Nothing |
| Works in web and phone apps | No | Yes, once the connector is added to your account |
| Updates | You do them | The provider does them |
| Where your data goes | Stays on your machine | Travels to the provider |
| Works offline | Yes | No |
💡 Rule of thumb: if the setup instructions mention
npx,uvxordocker run, it is a local server. If they hand you a URL, it is remote.
These are the endpoints that answered an MCP initialize request with no credentials on 6 October 2026. None asked for a login or a token, and all five are built around searching and reading, which is the safest category to start with.
| Server | URL | Best for |
|---|---|---|
| DeepWiki | https://mcp.deepwiki.com/mcp | Questions about public GitHub repositories |
| Cloudflare Docs | https://docs.mcp.cloudflare.com/mcp | Searching Cloudflare product documentation |
| AWS Knowledge | https://knowledge-mcp.global.api.aws | AWS docs and regional availability |
| Context7 | https://mcp.context7.com/mcp | Current documentation for code libraries |
| Hugging Face | https://huggingface.co/mcp | Finding models, datasets and papers |
DeepWiki turns a public repository into a browsable wiki and exposes it through three tools: read_wiki_structure, read_wiki_contents and ask_wiki_question. I called read_wiki_structure on the MCP specification repository and got back a numbered outline with sections on the transport layer and OAuth 2.1 authorization.
A prompt that works well: "Using DeepWiki, outline how the transport layer is organized in modelcontextprotocol/modelcontextprotocol." The assistant picks the right tool and returns the structure instead of guessing from memory.
Cloudflare's documentation server exposes search_cloudflare_documentation. Ask about Workers limits, R2 behavior or Pages settings and the answer comes from the live docs instead of whatever the model remembers.
It also suits anyone planning to host a remote MCP server of their own, since Workers is a common place to run one.
AWS publishes a server with aws___search_documentation, aws___read_documentation, aws___list_regions, aws___get_regional_availability and aws___retrieve_skill. The regional availability tool is the one that saves real time. Instead of opening five console pages to check whether a service runs in a given region, ask the assistant and let it call the tool.
Notice that this URL has no path at all. Not every server uses /mcp, so always copy the address exactly as the provider prints it.
Context7 offers two tools: resolve-library-id and query-docs. The first maps a library name to an identifier and the second pulls current documentation for it. That fixes the classic problem of a model writing code against an API that changed two releases ago.
Hugging Face answered the handshake at https://huggingface.co/mcp without credentials. It is built for finding models, datasets, Spaces and papers. Anything tied to your own account will need you to sign in, so treat anonymous use as the baseline and add a login only when you have a reason to.
💡 A server that failed my test: Semgrep's server is listed as open on older directories under an
/sseaddress. On 6 October 2026 that path returned404, and the/mcppath answered401 Authentication required. Directories age faster than endpoints.
Claude supports remote servers through custom connectors. Both Streamable HTTP and SSE work today, though SSE support may be dropped in a coming release, which is one more reason to choose /mcp URLs.
https://mcp.deepwiki.com/mcp.Connectors you add on the web also show up in Claude's phone apps, which is the biggest practical win over local servers: the same tools follow you to the couch.
💡 Menu names move. Anthropic renames settings pages between releases. If the labels above are off, search the connectors page for "custom".
On shared plans the process has two stages. An organization Owner adds the connector under Organization settings, then Connectors. Each member then opens their own connectors page, finds the connector in the list and clicks Connect. Members cannot skip the first stage, so ask your admin to add the URL once.
If you want to compare answers across chat models before you commit to one, PicassoIA lists Claude Sonnet 5, GPT 5.4 and Gemini 3.5 Flash in its large language models collection.
ChatGPT reaches MCP servers through custom connectors too, but only the remote kind. The server must be reachable over HTTPS from the public internet, and a local stdio server will never work here.
Custom connectors for general tools sit behind developer mode, and developer mode sits on paid plans. Sources disagree on exactly which tiers include it, and OpenAI has moved the line before. Check the switch in your own account before you plan around it.
Open Settings, then Connectors. The developer mode switch is near the bottom of that page. Turn it on and a create option for custom connectors appears.
💡 Handle with care: developer mode lets the model take write actions. For documentation servers that is harmless. For a server connected to your repositories or database, read each tool call before approving it.
| Step | Claude | ChatGPT |
|---|---|---|
| Where | Customize, then Connectors | Settings, then Connectors |
| Gate | Plan with custom connectors | Developer mode on a paid plan |
| Open servers | Leave OAuth fields empty | Choose No authentication |
| Phone use | Added on the web, works in the app | Depends on your plan and app version |
Free lists go stale and some entries were never real. You can test any URL in under a minute from a terminal.
This is the exact request I used. Swap in the URL you want to check.
curl -s -X POST https://mcp.deepwiki.com/mcp \
-H "Content-Type: application/json" \
-H "Accept: application/json, text/event-stream" \
-d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-03-26","capabilities":{},"clientInfo":{"name":"probe","version":"1.0"}}}'
A healthy server replies with a serverInfo block that names it. To see the tools it offers, send the same request with "method":"tools/list" and "params":{}.
| Response | Meaning | What to do |
|---|---|---|
| 200 with serverInfo | Open and working | Add it as a connector |
| 401 | The server needs a login | Use it only if you trust the provider |
| 404 | Wrong path or retired endpoint | Check the provider's current docs |
| Timeout | Server is down or blocking you | Try later, or skip it |
Three things break lists. Providers move from /sse to /mcp. Open servers add authentication once traffic grows. And some entries are marketing pages rather than endpoints. A directory entry is a lead, not proof. The handshake above is proof, and it only takes one command.
A connector is a door into your chat. Free does not mean unwatched, so a few habits matter more than any feature list.
A documentation server cannot delete a file. A server wired to GitHub or a database can. Start with read-only servers and add write access one tool at a time.
Watch for prompt injection as well. Text returned by a tool is text the model reads, and a hostile page or repository README can contain instructions aimed at the assistant. DeepWiki indexes public repositories, so its output is only as trustworthy as the code behind it. If the assistant suddenly proposes something unrelated to your question after a tool call, stop and read the raw tool output.
A remote server sees whatever the assistant sends it: your question, repository names, search strings. Four rules handle most of the risk:
Documentation servers rarely publish how long they log tool calls, which is one more reason to keep sensitive questions out of a chat that has them switched on.
Free public endpoints come with no service level, so expect friction.
aws___. A saved prompt that names a tool can break after an update.
| Problem | Likely cause | Fix |
|---|---|---|
| Connector will not add | URL is an /sse path or has a typo | Use the /mcp address |
| Asks you to sign in | Server needs OAuth | Sign in, or pick another server |
| Empty tool list | Server rejected anonymous access | Run tools/list and read the reply |
| Slow or failing answers | Rate limit on a free endpoint | Wait, retry, narrow the question |
| Wrong library docs | Ambiguous library name | Name the exact library and version |
Remote servers handle the reading. For the making, the same one-URL idea exists on PicassoIA. Image and video generation are available through a Replicate-style API, where you create a prediction and then poll for the result, and through MCP connections managed from your account. An account runs up to five predictions at once, so a batch of ideas finishes faster than a single queue would.
Start with PicassoIA Image for text-to-image. When you want to change a picture you already have, open PicassoIA Image Editor Pro and describe the edit in plain words.
A quick experiment that takes ten minutes: connect DeepWiki to your chat, ask it to summarize how a repository you like is organized, then paste that summary into Picasso IA as a prompt for an illustration of the idea. You will have tested a remote server, a chat model and an image model in one sitting.
Browse every image, video and language model at picassoia.com/en/all-models, pick one, and generate your first image today.
