Quick answer
To add a custom MCP connector to Claude safely, use only the official server address from a provider you trust, add it under Customize → Connectors → Add custom connector, sign in, then review which tools it offers and set the ones that spend money or change things to need your approval. Run a small test, watch your credits for the first few days, and remove the connector (in Claude and in the provider’s account) when you stop using it.
Key takeaways
- A connector gives Claude new powers, so treat it like installing an app with access to your account.
- Only use the official MCP address from the provider’s own website or documentation.
- Keep spending and publishing tools on “needs approval”, at least at first.
- Test with a tiny job and check the credit cost before anything big.
- Disconnect in two places: in Claude and in the provider’s account settings.
In this guide
- What is a custom MCP connector?
- What should you check before you connect one?
- Step 1: Get the server address from the provider
- Step 2: Add the connector in Claude
- Step 3: Sign in and review what you are granting
- Step 4: Set tool permissions
- Step 5: Run a small test
- Step 6: Keep an eye on credits and data
- How do you remove a connector cleanly?
- What mistakes should you avoid?
- Frequently asked questions
- The bottom line
- Related posts
This guide follows on from our article Build a Game With Claude Fable 5.2 and Higgsfield MCP: What One Chat Really Did, where WorldofAI connected the Higgsfield MCP to Claude and let it generate art, voices, a 3D model and a trailer in the middle of a chat. That set-up took the creator under a minute. Doing it safely takes a few minutes more.
📘 Main article: See the whole game build, the four-round method and the costs shown in Build a Game With Claude Fable 5.2 and Higgsfield MCP: What One Chat Really Did.
This guide answers one question: how do you add a custom MCP connector to Claude without handing over more access, data or money than you meant to?
What is a custom MCP connector?
It is a link between Claude and an outside service that lets Claude use that service’s tools during a chat. MCP stands for Model Context Protocol, an open standard for connecting AI assistants to tools and data.
Claude includes a directory of ready-made connectors. A custom connector is one you add yourself by pasting a server address, such as the one Higgsfield publishes for its image and video models. Once it is connected, Claude can call those tools, for example “generate an image” or “make a video”, and the results come back into your chat.
💡 Tip: A connector can usually do anything your account on that service can do. If your account can spend credits, publish files or read private data, assume the connector can too.
What should you check before you connect one?
| Check | Why it matters | Where to look |
|---|---|---|
| Is the provider trustworthy? | A connector can see what you send it and return content into your chat | Company website, reviews, how long they have been around |
| Is this the official address? | Fake or copied servers can collect your data | The provider’s own docs or MCP page, not a forum post |
| Which tools does it offer? | Some tools only read; others spend money or publish | The provider’s docs, then Claude’s connector settings |
| How is it paid for? | Generations often use credits or a paid plan | Pricing page and your account balance |
| What happens to your data? | Prompts and files may be stored or used for training | Privacy policy and terms |
| Can you use the output commercially? | Matters for anything you sell or publish | Terms of service |
Your Claude plan matters too. Custom connectors are not available in the same way on every plan, and on business plans an administrator may need to switch them on. Check Anthropic’s help pages for your plan before you start.
Step 1: Get the server address from the provider
Copy the MCP server address from the provider’s official page. It is a web address, often ending in /mcp. For example, Higgsfield publishes its address on its MCP page and in its documentation.
Type the provider’s website address yourself rather than following a link in a comment or message, and check the domain matches the company’s main site.
Step 2: Add the connector in Claude
- Open Claude and go to Customize (labels can change slightly between versions).
- Choose Connectors, then the + button.
- Pick Add custom connector.
- Give it a clear name, such as the provider’s name.
- Paste the server address and confirm.
Claude may show a warning that custom connectors are not verified by Anthropic. That is normal, and it is a good reminder of why the checks above matter.
Step 3: Sign in and review what you are granting
Most connectors ask you to sign in to the provider’s account in a pop-up window. This uses a standard sign-in flow, so you do not paste a password or API key into Claude. WorldofAI’s Higgsfield set-up worked this way.
Read the permission screen before you approve it. If it asks for far more than the tool needs, for example access to your email for an image generator, stop and ask why.
⚠️ Warning: Never paste API keys, passwords or recovery codes into a chat to “help” a connector work. If a set-up guide asks you to, use a different provider or the official instructions.
Step 4: Set tool permissions
Decide which tools Claude can use freely and which ones need your say-so each time. In Claude’s connector settings you can usually see the list of tools and choose how each is allowed.
| Type of tool | Example | Suggested setting |
|---|---|---|
| Read-only | List models, check status, search a library | Allow |
| Creates something that costs credits | Generate an image, video or voice line | Needs approval |
| Publishes or shares | Host a page, post to a site | Needs approval |
| Deletes or changes account settings | Delete files, change plan | Needs approval, or do it yourself |
In a long unattended run, like the hour-long game build in the video, “always allow” on a credit-spending tool can use up a balance faster than you expect. For extra protection, read How to Add a Kill Switch to Your AI Agents and Automations.
Step 5: Run a small test
Start a new chat and give Claude one small job that uses the connector. For an image tool, ask for one image at the smallest size. Then check three things:
- Did Claude ask before using the tool? If it did not, revisit your permission settings.
- What did it cost? Compare your credit balance before and after. In the WorldofAI video, one key-art image cost 2.75 Higgsfield credits.
- Where did the output go? Find the file in the provider’s account so you know where your assets are stored.
Do this now: before your first big project, write down the credit cost of one test job. Multiply it by the number of assets you plan to make, and you have a rough budget.
Step 6: Keep an eye on credits and data
- Check your balance after each session for the first week.
- Set a spending cap or low-balance alert in the provider’s account if it offers one.
- Keep confidential material out of prompts sent to third-party tools unless their terms allow it and your business is happy with that.
- Watch for odd instructions in tool results. Content returned by an outside service is data, not orders; if Claude suddenly wants to do something unrelated, stop the chat.
The same thinking applies to any assistant with access to your accounts. Our guide How to Let an AI Assistant Read Your Email Without Sending Anything uses the same “read freely, act with approval” approach for email, and AI Agent Internet Access: The Three Zones Your Agent Will Meet explains why outside content needs care.
How do you remove a connector cleanly?
- In Claude, go to Customize → Connectors, open the connector and choose to disconnect or remove it.
- In the provider’s account, look for connected apps, integrations or API access and revoke Claude’s access there too.
- Download any assets you want to keep, then cancel the plan if you no longer need it.
Removing it in both places matters. Disconnecting in Claude stops Claude using it; revoking in the provider’s account stops the sign-in token working at all.
What mistakes should you avoid?
- Using an address from a comment or video description without checking it against the provider’s own site.
- Leaving every tool on “always allow” for convenience.
- Skipping the test run and discovering the cost after a big job.
- Forgetting connectors you no longer use. Review your list every few months.
- Assuming generated files are yours to sell. Check the commercial terms first.
Frequently asked questions
Do I need an API key to add a custom MCP connector?
Often not. Many providers, including Higgsfield, use a normal sign-in pop-up. If a provider asks you to paste a key into a chat, be cautious.
Is a custom connector safe?
It is as safe as the provider behind it and the permissions you give it. Use official addresses, keep costly tools on approval and remove connectors you stop using.
Can I use custom connectors on any Claude plan?
Availability varies by plan and can change. Check Anthropic’s help pages, and ask your administrator on a business plan.
What is the difference between a connector and a plugin?
In everyday use they do the same job: they give the assistant extra tools. “Connector” is the name Claude uses for MCP-based integrations.
Will the connector see my other chats?
It receives what Claude sends it when it uses a tool, not your whole history. Still, avoid putting confidential details into requests that go to outside tools.
How do I know how much a connector is spending?
Check the balance or usage page in the provider’s account. Run a small test first so you know the cost of one job.
The bottom line
Custom MCP connectors are what turn Claude from a chat window into a small production team, as WorldofAI’s game build shows. They are quick to add, but they act with your account’s power. Use the official address, keep spending and publishing tools on approval, test with one small job and remove what you do not use.
Once your connector is set up, head back to Build a Game With Claude Fable 5.2 and Higgsfield MCP: What One Chat Really Did for the four-round method to put it to work.
Related posts
- Build a Game With Claude Fable 5.2 and Higgsfield MCP: What One Chat Really Did
- How to Turn One Reference Image Into a Visual Style Brief for AI
- How to Add a Kill Switch to Your AI Agents and Automations

