How to Connect Your AI Assistant to Feedpik

Feedpik is built so your own AI assistant can run it: pick your reading, tune your ranking, add sources, keep a list waiting for you every morning. But "connect your agent" turns out to mean something different in each product, and one of them cannot do it at all yet.

This is the whole map. Find your assistant, follow its section, skip the rest.

First, which kind of assistant do you have?

Every route below is one of three, and which ones are open to you depends on a single question: can your assistant run commands on a computer?

	Can run commands	Route
Claude Code, Codex, Gemini CLI, OpenClaw	Yes	The CLI, or a local MCP server
ChatGPT web, Claude web, Gemini Enterprise	No	A remote MCP connector
Gemini consumer app	No	Not possible yet — see below

Assistants that run in a sandbox — which is most of the ones you talk to in a browser — cannot execute npx or anything else. That is a deliberate safety boundary in those products, not something to work around. For them the remote connector exists.

Everything connects to the same account and gets the same abilities either way.

---

Claude Code

The simplest of the lot, because it has a terminal.

Tell it:

Set up Feedpik for me. Run npx feedpik skill to read the instructions, then get me connected — I'll approve it in my browser.

It will run npx feedpik@latest login --no-wait, which prints a short code and exits. It hands you the code, you open feedpik.com/cli-auth and approve it while signed in, you tell Claude it is done, and it claims the token. You never type a command yourself.

Two details make this work that are worth knowing:

npx feedpik skill prints the current agent instructions to the terminal. That matters because web-fetching tools in most assistants pass a page through a summariser and return a description of it rather than the text — so an agent that "reads" a documentation URL often gets a paraphrase. Printing to stdout gives it the real thing.

The --no-wait flag exists because the normal login waits for approval, which would leave the command hanging. This variant prints and exits, so the agent stays responsive while you approve.

If you would rather have structured tools than shell commands, there is also a local MCP server:

claude mcp add feedpik -- npx -y -p feedpik feedpik-mcp

Restart Claude Code afterwards — MCP servers are connected when a session starts. The -p feedpik is not optional: the command is feedpik-mcp but the package is feedpik, and without it npm looks for a package that does not exist.

Claude (web and desktop)

Claude in a browser cannot run commands, so it uses a remote connector. You add this, not Claude — it has no way to install anything into your account.

Settings → Connectors → Add custom connector, and paste:

https://feedpik.com/api/mcp

Approve it in the browser window that opens. No token is ever created, shown, or pasted anywhere.

Then, in a conversation:

Connect to my Feedpik account and tell me when you're in. Don't try to run any npx command — you're sandboxed and it will be refused. My account is new, so set it up and tell me what you change.

The last sentence saves a round trip. Without it, an assistant reading our documentation may still try the CLI first, get blocked by its own permission layer, and stop.

ChatGPT (web)

Same shape as Claude, with one significant caveat: custom MCP connectors require developer mode, which is only available on ChatGPT Business, Enterprise and Edu. On Free and Plus, this route is closed today. That is OpenAI's gating, not ours, and nothing on our side changes it.

If you are on an eligible plan: Settings → Apps & Connectors → enable developer mode, then add a custom connector pointing at:

https://feedpik.com/api/mcp

Workspace admins may need to allow it first, under Permissions & Roles.

Then talk to it exactly as you would to Claude above.

Codex

Codex has a shell, so treat it like Claude Code — the CLI route works unchanged:

Run npx feedpik skill, then connect me to Feedpik. I'll approve it in my browser.

Codex also speaks MCP if you would rather register the server, but for reading and tuning a feed the CLI is fewer moving parts.

Gemini CLI

Add Feedpik as an MCP server in your Gemini CLI settings — the same local stdio server Claude Code uses:

{
  "mcpServers": {
    "feedpik": {
      "command": "npx",
      "args": ["-y", "-p", "feedpik", "feedpik-mcp"]
    }
  }
}

The server drives its own login: ask Gemini to connect, it returns a code, you approve it in the browser.

One caution: Google is migrating Gemini CLI's extension system to Antigravity CLI, so the surrounding configuration may move. The MCP server itself does not change.

Gemini (consumer web app)

Not possible today. The consumer Gemini app's connectors are partnership-only — there is no way for you to add a custom MCP server to it, whatever the URL. This is a limitation of that product, and we will update this post when it changes.

Two things that do work:

Gemini Enterprise can connect any Streamable HTTP MCP server, so https://feedpik.com/api/mcp works there. Note it does not support the legacy SSE transport, only Streamable HTTP — which is what our endpoint speaks.

Gemini CLI works, as above.

OpenClaw

OpenClaw ships stdio and HTTP MCP transports, so either route works. Local, in openclaw.json:

{
  "mcpServers": {
    "feedpik": {
      "command": "npx",
      "args": ["-y", "-p", "feedpik", "feedpik-mcp"]
    }
  }
}

Or point it at the remote endpoint, https://feedpik.com/api/mcp, if you would rather not run anything locally.

Hermes, and anything else

Any agent that can make HTTP requests can use Feedpik. There is nothing vendor-specific anywhere in this — the products above differ only in how they let you register a server.

• If it has a shell: npx feedpik@latest login --no-wait, then --claim <code> after approval.
• If it speaks MCP: either https://feedpik.com/api/mcp (remote, OAuth) or the feedpik-mcp stdio server.
• If it only does HTTP: create a key at feedpik.com/tokens and send it as Authorization: Bearer <token>. Full reference at feedpik.com/skill.md.

Ask for the full scope set — an agent that can only read cannot tune anything, and a narrower key sends you back through approval the first time it needs more.

---

What happens the first time it connects

On a new account, an assistant that has read our instructions will:

1. Propose a reading profile — interests, topics to mute, languages — drafted from what it already knows about you, and show it in plain language for you to correct. It should not interview you with a list of questions.
2. Ask before adding sources. Subscriptions change what arrives every day, so this one always gets a question.
3. Fill your reader. It publishes around 20 articles into the Agent Picked view — a page you can open tomorrow, on your phone, without the assistant.
4. Offer to keep it filled, on a schedule you choose.

Then tell you what it changed. If it says something as vague as "your feed will emphasise AI", ask for specifics — our instructions require it to name what it muted and which way each weight moved, and to give exact values to anyone who asks. Your algorithm is not much use to you if you cannot find out what happened to it.

When it does not work

"The command was blocked" — your assistant is sandboxed. Use the remote connector instead; that is what it is for.

The assistant read our docs and got them slightly wrong — likely its fetch tool summarised the page. Ask it to use the MCP connection, or to run npx feedpik skill if it has a shell. Both deliver the real text.

Adding the connector fails — send us the error. Connector setup involves an OAuth discovery handshake, and a client that is strict about one part of that specification can fail in ways nothing in the interface explains.

Stuck somewhere else? Reply to this post or open an issue — connection problems are worth a specific answer, because almost always the fix is one line and finding it is the hard part.