Guide
Integrations in Fibery are quite unusual. They replicate a part of an external app domain and feed data into Fibery to create several interconnected Databases.
A couple of dozens of such services are already available as Integration templates in Fibery. However, sometimes you may need a custom integration which has not yet been added to Fibery. In this case, to speed up development, we encourage to use coding agents in a way described below. In the best case you can get working integration from the first try (this was the case for Stripe, for example).
Proposed solution focuses on continuous one-way sync (data from External App to Fibery).
The flow is relatively simple. You clone repository with an integration template and use prompts we've made for you to transform this template into a working integration. Then you can share and test it in your Fibery workspace, and finally deploy to Vercel and use on production when you are happy with the results.
https://www.loom.com/share/76c02d537c4c46038ed4f5b84b403140
Pre-requisites
Install on your machine:
Setup
Getting repository & external-app specific information ready to develop integration:
Login into your GitHub account
Create a new repository using Fibery connector-template. It is a ready-to-use integration template with boilerplate code to ease the life of AI coding agent.
git clone newly created repo locally.
Research
Go to your favourite AI provider (Anthropic, OpenAI, etc.) and enable Research mode (we used Sonnet 4.5 without Extended thinking).
Copy prompts/integration-research-agent-prompt.md prompt from cloned repository into the Research agent input field. It is a baseline prompt for the research process. Then, in your message, add documentation link and integration requirements if you have any. That's how it looks for Stripe:
Wait for the research to finish. It should generate a comprehensive report of your external app's API which will help enormously AI coding agent later to develop integration quickly and efficiently.
Develop
There is a chance that you will not have to develop anything. Instead, you will form a prompt and, hopefully, AI agent will handle the development for you:
Open your favourite AI coding agent tool in cloned repository (in our case, it was Claude Code with Opus 4.5)
Form the initial prompt:
Copy-paste prompts/integrations-system.md prompt to the Agent's input field. Type \ + Enter to create a new line.
Copy-paste the report document you've got on Research step. Claude console should look like this:
Put it in auto-accept mode if possible (Shift + Tab for Claude Code).
Go get some ☕. Hopefully, after Claude is finished, you will get the basic version of integration working.
Test
After the agent is done coding, you can test your locally hosted integration right away in your Fibery workspace. It is possible by using ngrok:
Run pnpm start to start integration service (notice the port it uses, you will use it in the next step)
While pnpm server is running, open another terminal window and run ngrok http <your-port> command — it will open your app to the web temporarily under url like this https://xxxxx.ngrok-free.dev.
Add your integration as custom integration to your Fibery workspace (new Space → Sync from → Add custom integration) and use the url from ngrok:
If there are errors, copy them into Claude Code and ask to fix.
Deploy
We recommend to use Vercel for a quick deployment. It has a free plan with bandwidth of 100GB/month, which should be enough for testing how your integration behaves when deployed. The repository you cloned is already pre-configured to work with Vercel, too:
Create a new free account here
Install vercel CLI by running brew install vercel-cli for MAC or pnpm i -g vercel for any machine
Run vercel login
To test how it works locally, run vercel dev
When ready, run vercel deploy and accept defaults for every question
Lastly, add publicly-available link as custom integration in Fibery and start syncing the data!
FAQ
How long can I iterate on a solution until I run out of context?
Context window is usually long enough for a couple of iterations on integration (e.g., to get the basic version working, and a couple of new objects on top). However, it is not a big problem since most modern coding agents (including Claude Code) have automatic compacting of context and should be able to carry on even after you run out of initial context window.