Codend launches AI agent to automate code upgrades and pull requests

From description to pull request

Codend is an AI coding agent designed to turn high-level descriptions into functional software updates. Instead of manually editing files, developers can provide instructions in plain language to handle complex tasks. For example, a user can ask the agent to upgrade a project to a specific version of a framework like Next.js or React.

Once a task is described, the agent takes over the heavy lifting of the development cycle. It plans the necessary changes, writes the code, and runs the relevant tests to ensure everything remains functional. As codend.dev reported, the goal is for the user to focus on making decisions while the agent handles the execution of those choices.

The final output of this process is a pull request, which is a request to merge new code into a main project. This allows the human developer to see exactly what the AI did before anything is officially changed. The agent handles the technical steps of creating a new branch, editing files, and running build commands automatically.

Security and sandbox isolation

A primary concern for developers using AI tools is the security of their local environment and private data. Codend addresses this by running every single task in its own separate Linux sandbox. This means the agent's files and commands are isolated and never touch the user's personal computer or other active tasks.

The tool also employs a strict permission model regarding the repositories it can access. The GitHub App only reaches the specific repositories that a user chooses to install it on. It uses short-lived tokens for each individual call, which limits the potential surface area for unauthorized access.

Furthermore, the system is designed to keep private information out of shared environments. Shared chats are configured to exclude results from connected apps and any files the user has attached. Users can also disconnect any connected app at any time through a one-click connector interface.

Human-in-the-loop workflow

Despite the automation, Codend is built so that the human user stays in complete control of the project. Nothing that matters happens without an explicit confirmation from the owner. The agent waits for the user's okay in a chat interface before it performs actions like opening an issue, commenting, or creating a pull request.

The workflow ensures that the main branch of a project remains untouched by the AI. Codend works on its own branch and only moves toward the base branch after a human review. This structure prevents the AI from accidentally breaking the production version of a website or application during its work cycle.

If a user wants to change something the agent is currently working on, they can simply reply in the task description. The agent will then update the same branch and the existing pull request. This iterative process allows for a collaborative experience where the human provides the direction and the agent provides the labor.

Integrations and automation

Codend connects with a variety of popular developer and productivity tools to provide context for its tasks. It can read data from platforms like Notion, Supabase, and Google Drive to understand the requirements of a project. It also integrates with communication and deployment tools like Slack and Vercel.

One of the key features for teams is the ability to run tasks on a schedule. This is particularly useful for recurring maintenance work, such as nightly checks or regular dependency updates. Developers can hand off these repetitive chores to the agent and review the results whenever they are ready.

The platform offers a live preview feature to see changes before they are finalized. This allows for a more visual confirmation of the work performed by the agent. By combining these integrations, Codend aims to act as a single agent that can plan, write, test, and open pull requests for a team.

Onboarding and availability

New users can begin exploring the tool with a set of welcome credits. These are supplemented by daily credits provided during the first 7 days of use. This allows developers to test the agent's capabilities on various tasks before committing to a paid plan.

The pricing model is structured so that users can upgrade their limits as their needs grow. This makes it accessible for individuals who want to try out the automation on a small scale or for larger teams with higher volume requirements. The platform is designed to scale as a team moves more of their maintenance work to AI agents.

The tool is currently available for those looking to automate the more tedious parts of software engineering. By focusing on the transition from plain language to merged code, Codend seeks to streamline the development lifecycle. It provides a way for developers to manage complex updates without having to manually navigate every line of a dependency change.