How to Connect Backrow to Claude or ChatGPT with MCP

You want your AI assistant to read your notes, transcribe your recordings, and pull your flashcards — without copy-pasting between apps. That's exactly what backrow's MCP server is for. This is a step-by-step guide to connecting it to Claude, ChatGPT, Cursor, or any other MCP host.

It takes about five minutes and four lines of config.

What you'll need

• A backrow Pro ($44.99) or Max ($89.99) plan. The API and MCP server are a paid-plan feature.
• An MCP-capable host. The server works with Claude Desktop, Cursor, ChatGPT, VS Code, and Windsurf — any host that supports remote Streamable-HTTP MCP servers.
• A bk_live_ API key.

Step 1: Create an API key

Sign in to backrow, go to the developer area, and create a new API key. It will start with bk_live_. Copy it somewhere safe — treat it like a password, because it can spend your credits.

[VERIFY: source docs reference key creation at both backrow.ai/developers and backrow.ai/platform — confirm the canonical URL for the key-management screen before publishing this step.]

While you're there, note the scopes on the key. backrow keys are scope-gated, so the tools your agent can see and call depend on what the key is allowed to do.

Step 2: Add the server to your host

The server lives at https://backrow.ai/mcp and speaks Streamable HTTP. You connect with a bearer token in the Authorization header. The config is the same shape across hosts:

{
  "mcpServers": {
    "backrow": {
      "url": "https://backrow.ai/mcp",
      "headers": {
        "Authorization": "Bearer bk_live_..."
      }
    }
  }
}

Replace bk_live_... with your real key.

Claude Desktop

Add the backrow block to your mcpServers configuration and restart the app. Claude Desktop supports remote MCP servers, so the URL-based config above is what you want. [VERIFY: confirm the exact config file path and whether the current Claude Desktop build accepts a remote Streamable-HTTP URL directly, or whether it needs an mcp-remote bridge for this host version.]

Cursor

Add the same backrow entry to Cursor's MCP settings and reload. [VERIFY: confirm the current Cursor MCP settings location and remote-server field names.]

ChatGPT, VS Code, Windsurf

Each host has its own settings screen for adding an MCP server, but the three values are always the same: the URL https://backrow.ai/mcp, the transport (Streamable HTTP), and the Authorization: Bearer header with your key. [VERIFY: a handful of older host versions still need an mcp-remote or SSE shim for remote servers — note this only if your target host requires one.]

Step 3: Verify the connection

Before trusting it, prove the server is reachable. A no-auth GET returns the server's transport, protocol version, and live tool count:

curl https://backrow.ai/mcp

You should see something like:

{
  "protocol": "mcp",
  "protocolVersion": "2025-06-18",
  "transport": "streamable-http",
  "server": { "name": "backrow", "title": "Backrow", "version": "1.0.0" },
  "toolCount": 98,
  "docs": "https://backrow.ai/developers#mcp"
}

If that responds, the server is up. Once your host is configured with a valid key, ask your agent to list backrow's tools — it should return the set your key's scopes allow.

Step 4: Try a first command

Now use it. A simple first request to confirm everything works end to end:

"Using backrow, list my recent notes and tell me how many credits I have left."

The agent calls the relevant tools and reports back. If it can read your notes and your credit balance, you're connected. From there you can build up to real workflows — transcribe a recording, generate flashcards, take a quiz.

Troubleshooting

• 401 / auth errors. The key is missing, malformed, or on a Free/Starter/Scholar plan. The API and MCP server require Pro or Max. Confirm the header is exactly Authorization: Bearer bk_live_....
• The agent sees no tools. tools/list only returns tools your key's scopes allow. Check the key's scopes.
• insufficient_credits errors. You're out of credits. There's one balance across the app, API, and MCP; top up or wait for your monthly refresh. Failed calls refund automatically, so a single error won't have charged you.
• Rate-limit errors. The limit is 60 requests per minute per key. Slow the agent down or batch fewer calls.
• Host won't accept the URL. A few older host builds need an mcp-remote bridge for remote Streamable-HTTP servers. [VERIFY: confirm whether your specific host version needs the shim.]

A note on safety

One bearer key can spend your credits, so guard it like a password. The good news: the balance is a hard cap, so a misbehaving agent can't run up an unbounded bill — it stops at insufficient_credits. Scopes limit what any key can touch, and every credit transaction is auditable through the API and MCP server.

A few practical habits keep a key safe. Don't paste it into a shared chat or commit it to a repository. Use a separate key for anything experimental so you can revoke it without breaking your main setup. And glance at your credit-transaction history after wiring up a new automation — if something is calling tools you didn't expect, you'll see it there before it matters.

Why this is worth doing

The payoff of connecting your assistant to backrow isn't novelty. It's that the friction between "I have a question about my own material" and "I have the answer" drops to a sentence. No opening the app, no hunting through notes, no copy-pasting a transcript into a chat window. You ask your assistant, it calls the right tool over MCP, and it answers from your actual recordings, notes, and learning data.

That compounds. The more of your learning lives behind one connected interface, the more your assistant becomes genuinely useful instead of generically helpful — it answers from what you've captured, not from the open internet. Setting it up once, with four lines of config, is what makes that possible.

Next steps

That's the whole setup: a Pro/Max plan, a bk_live_ key, four lines of config, and a verify call. The full host-specific instructions and tool reference live at backrow.ai/developers.