How to connect fal MCP to Claude Code or Cursor

Connect fal AI MCP by adding https://mcp.fal.ai/mcp-relay in a compatible assistant, then complete browser sign-in. For Claude Code, use the documented command below; for Cursor, install the fal Marketplace plugin and sign in from Customize. The official server guide↗ (opens in a new tab) distinguishes this generation connection from its other servers. fenbs can keep your tested briefs and accepted findings outside the chat so the next assistant starts with the same work.

Which fal server do you need?

MCP is a standard way for an AI assistant to use an outside tool. For this workflow, the assistant is the client and fal supplies the generation server. You need a compatible client, a fal account, and generation credits; organization rules can also affect whether a connection is available.

Use the relay address above for media work. The documentation server at https://fal.ai/docs/mcp searches public documentation and cannot generate media. Platform MCP is a separate connection for account and Serverless operations. A successful connection to the wrong server will not provide the generation workflow you wanted.

fal’s generation connection uses OAuth, meaning you authorize it through browser sign-in, and Streamable HTTP, its supported remote connection method. This setup does not require pasting an API key. If you need calls from your own application instead, use the separate fal API guide.

How do you connect Claude Code?

Follow fal’s Claude Code setup↗ (opens in a new tab). This is the Claude Code route; do not assume the command configures every Claude interface. Run it in your terminal, then finish authentication inside Claude Code.

claude mcp add --transport http --scope user fal https://mcp.fal.ai/mcp-relay

1. Run the command above to add the remote server.
2. Open Claude Code and enter /mcp.
3. Select fal, then choose Authenticate.
4. Complete browser sign-in and return to Claude Code.
5. Ask the assistant to find a fal model without running it. Confirm you get model results before requesting media.

Keep setup verification separate from your first generation. Finding a model checks the connection without asking for a finished asset. Do not paste a credential into chat to work around an incomplete sign-in; return to the documented authentication step instead.

How do you connect Cursor?

The official Cursor guide↗ (opens in a new tab) gives a Marketplace route and a manual configuration route. Start with the Marketplace option if you want the documented plugin setup. Each user signs in to authorize their own connection.

1. Open fal’s listing in the Cursor Marketplace and add the plugin.
2. Open Customize in Cursor.
3. Find fal in the MCP server list and complete browser sign-in.
4. Open an Agent chat and ask Cursor to find a fal model without running it.

For manual setup, the same guide documents ~/.cursor/mcp.json. Preserve any existing connections while adding the fal entry. Follow the guide for that file’s full structure instead of replacing your configuration with an example that removes other servers.

The shell installer described in fal’s overview is for macOS and Linux. It is not a Windows setup command. Use the client-specific instructions above when those are the route you intend to follow, and record which route you actually tested.

What should your first generation request say?

Begin with a small, easy-to-review brief. Say what the output is for, the subject, what must remain consistent, and what you want back. Ask the assistant to identify the model and proposed inputs before generating, so you can catch a misunderstanding while it is still a plan.

Find a fal model for a draft website image.
The subject is a plain ceramic cup on a light table.
Keep space on the right for a headline.
Show the chosen model and proposed inputs first.
Ask before generating. Afterward, return the result
and summarize the exact inputs and what needs review.

Treat “ask before generating” as an instruction to your assistant. fal explicitly distinguishes that prompt convention from a server-enforced spending approval gate. You still need to understand the account you connected and review the proposed action. An instruction in a chat should not be described as a technical control.

Compare the output with your requirements rather than asking only whether it looks good. For the cup example, check whether the subject is recognizable and the headline space is usable. Keep the accepted wording in a reusable record; building a prompt library is the adjacent job of organizing those instructions.

What goes wrong after the server is added?

• Only documentation answers appear: check that you added the generation relay rather than the documentation server.
• Authentication remains incomplete: repeat the client’s browser sign-in and authorization step.
• A colleague cannot use your setup: each user needs to authorize their own connection.
• The assistant suggests a different server: distinguish media generation from Platform MCP before proceeding.
• A result exists but the next session cannot continue: save your instructions, accepted inputs, and output location before ending the chat.

Separate connection problems from creative revisions. If model discovery works but the resulting image misses your requirements, capture the failed requirement and change one instruction at a time. Reinstalling a working connection does not explain why you rejected the asset.

For another service’s connector, see the Higgsfield MCP guide. Keep its setup separate from fal’s. The broader AI connections topic covers how to think about connections across work tools.

How to keep reusable fal briefs with fenbs

Keep the work so you can carry on from exactly where you left off, on any device and with another AI. Save your accepted instructions, the model used, the accepted output’s location, and the next requested change. Leave secrets out of that record.

1. Create a free account.
2. Connect Claude Code or Cursor to fenbs from Integrations. Keep the fal generation connection as its own connection.
3. Save the accepted brief in Skills and prompts. Add a task with the chosen model, output link, and review notes.
4. Before you stop, ask the assistant to save what is done and the next step. In a new chat, ask it to read that saved work before making another version.

fenbs keeps the shared work record and reusable prompts; it does not generate the media or import your fal chat automatically. A one-off generation may need only your current chat and saved file. The extra record is useful when you want another session to understand which result you kept and why.