Have you ever wished you had a code review assistant of your own that could instantly give feedback on each new Pull Request (PR)?
Manual code reviews are a necessary part of the development process, but they are expensive, slow, and inconsistent. As of writing this article, tools like GitHub Copilot's Code Review feature are premium, requiring subscriptions:
- Copilot Pro: ($10 USD/month or $100/year)
- Copilot Pro+: ($39 USD/month or $390/year)
Can we achieve this within a reduced budget? YES.
This article is designed to help you build your own persistent, automated AI Code Review Assistant hosted for free (or near-free) using a powerful, controlled workflow.
Tools
| Role in the System | Why We Use It | |
|---|---|---|
| n8n | Workflow Creation | An AI workflow automation tool that connects GitHub to OpenAI, manages the logic, and formats the output. |
| NHost | Managed Database | Provides a free-tier, managed PostgreSQL database where n8n stores all workflow data and credentials. |
| Render | Hosting n8n | Hosts our n8n web service, assigns a domain URL, and connects to the NHost database. |
| OpenAI | Review Code | Provides the LLM intelligence to critique the code. |
Why PostgreSQL (NHost) over SQLite?
n8n uses SQLite by default, which is great for local development but does not support highly concurrent transactions and struggles with scalability. We need a solution that can scale and adapt based on the number of Pull Requests being raised to the repository. NHost provides a free tier that fits well for a personal project while offering PostgreSQL's stability.
Database Setup and Hosting
I have followed the exact steps mentioned in this article → The Ultimate Guide to Self-Hosting n8n for Free using Render and Nhost - DEV Community to set up the underlying infrastructure.
Step 1
Create an account in NHost, start a project, and set up a Database. This will generate environment variables as mentioned in the post.
Step 2
Sign up for an account on Render, create a new Web Service, and choose n8nio/n8n:latest as the Docker image. Add the environment variables generated in NHost to the Render project.
This will host your n8n application on the cloud and generate a domain name.
Step 3
To ensure GitHub can reach your bot, you must correct the webhook URL inside n8n. Set the following environment variables in your Render Web Service settings:
-
WEBHOOK_URL:https://[your-render-domain].onrender.com/ -
N8N_HOST:[your-render-domain].onrender.com -
N8N_PROTOCOL:https
Credentials Required
GitHub Access Token: Generate a Personal Access Token from your GitHub settings and ensure it has the repo scope checked (required for reading and commenting).
OpenAI API Key: You'll need an active key. If you are a new user, you might have received free trial credits, but many users now need to pay the minimum of $5 to activate the key and avoid the insufficient_quota error.
Cost Control: I have used gpt-5-nano for this project as it consumes the fewest tokens. You can use any capable model you wish, but remember that the Mini models offer the best balance of power and cost.
Creating the Workflow
In N8n, you’ll find a list of actions in the right pane that you can drag and drop to the canvas. The actions needed for our workflow are Github Trigger, HttpRequest, Code in JS, OpenAI.
Github Trigger
In Github, generate an Access Token from the settings and provide all access.
In the right pane inside n8n, search “Github”, and drag and drop “on a pull request” trigger. This will add the same to your canvas as the first step. Click on the item to open the settings.
Once saved, the GitHub Trigger node will display a unique Webhook URL.
Fetching the Code Diff
The goal is to fetch the raw code changes from the PR event data. The previous node provides an object containing the PR details. You must construct the GitHub API URL to get the file changes. For simplicity, you can often grab the {{ $json.body.pull_request.url }}/files to get the list of files in the PR. This HTTP request needs to use the GitHub Access Token credential for authorization.
Defining the AI Persona
The next step is to define the prompt for OpenAI to start reviewing your PR. Inject reviewPrompt to define the AI's role, ensuring high-quality, structured feedback. Paste a detailed prompt into this node. The persona "Senior Software Engineer" is vital for quality and technical feedback. This node structures the prompt and the code diff into the format expected by the OpenAI Chat Model. Add your own customizations inside the instructions variable.
Generating the Review
This is the step where you integrate OpenAI into the workflow. Select the cheapest capable model, such as gpt-5-nano. In the node settings, map the variables from the previous Function node. Add the API key that was generated previously.
Drag and drop {{$json.reviewPrompt}} to the Prompt field.
Formatting the Output
This is the node labeled "Translate to Github version" in the diagram. The goal is to take the raw Markdown text generated by OpenAI and wrap it in the exact JSON payload required by the GitHub Comments API.
Posting the Comment
The final step of the workflow: Node labeled "Post comments on Github". The goal is to send the formatted review back to the Pull Request.
Details:
- Method: POST
- API Endpoint: Construct the URL using data from the trigger node: {{ $('Github Trigger').item.json.body.pull_request.url }}/reviews
- Body: Use the JSON output from the previous step - {{ $json.comments }}
- Authentication: Select the GitHub Access Token credential.
Output
The agent posted the review comment to the Pull Request based on the instructions provided. Additionally, you can provide a structure or template, the agent can follow while creating the comment.
Conclusion
This is a fairly basic prompt that could be improved by adding more instructions to the Code Review Prompt node.
To ensure your solution stays robust and economical, implement conditional skipping logic in an IF Node to skip the expensive OpenAI call for trivial or work-in-progress PRs, and add schema validation to your final Function Node to ensure the AI's JSON output is always clean and accepted by the GitHub API.
You are now in full command of your automation. Happy coding!
Top comments (0)