AI tools have advanced to the point where you don’t even need a coding editor to code anymore. You don’t even need to code to code nowadays. You just need to provide natural language prompts and tell the AI what you want to create.
I have been using OpenClaw to build my own AI coding workflow, which revolves around Claude Code, Discord and many other tools. It’s actually quite simple. Discord acts as an interface and I connect all the tools the agent needs. I can then ask the bot to fix a specific bug, create a new page for my website, or add a feature to my product, and it will do it. All I really need to do is check the result.
OpenClaw and Discord work very well together
You can even use Slack if you want.
The goal here is to build a always available AI agentand OpenClaw has many of the primitives you need. It officially supports Discord as a bot interface, including DMs and guild channels. You can also use something like Telegram or WhatsApp, but I prefer Discord because you can create separate channels for different projects. This prevents context from getting confused and also makes it easier for you to follow what’s happening.
You can also create separate threads, which makes conversations easier to track rather than having everything in one long chat, as is the case with WhatsApp. If you’re creating this for work, you can also connect it to something like Slack, Microsoft Teams, or any other device your organization uses.
Start by installing OpenClaw on your device. There’s a simple script you need to run, after which the rest is mainly a matter of following the on-screen instructions. OpenClaw installs all dependencies itself. You only need to select an AI model, among other things, during setup, and it will be prompted throughout the process. Please note that during initial setup, you can select any AI provider you want. It doesn’t have to be the same one you’ll eventually use to code. You can configure that later.
Once you are done with the initial setup, you will need to replace the default interface with Discord. Start by creating a dedicated private server on Discord, then create channels relevant to your workflow. For example, I have agents, features, and reviews. Now you need to create an agent for Discord. Discord already has built-in bot functionality, so go to the Discord Developer Portal and create a bot there.
Once you have created the bot, generate a bot token. This token is essentially like an API key, so keep it somewhere safe. Then invite the bot to your Discord server, select the required scopes, and grant permissions to view channels, read messages, and send messages. I would also enable reactions and messages in threads to help keep conversations clean and organized. Now get your Discord ID and give the Discord token to OpenClaw. Create a configuration file to enable Discord on OpenClaw, restart OpenClaw, and then return to Discord. Your agent should now be online.
Code Claude is for me
Once your Discord interface is up and running, the next step is to connect the tools that will actually do the coding. OpenClaw should act as the orchestrator here, not necessarily the tool that writes each line of code. I use Claude Code as the primary coding agent, but you can use Codex, OpenCode, Gemini CLI, or another supported coding harness instead. OpenClaw supports running external coding agents, such as Claude Code, through its ACP integration, so you can start and manage coding sessions while continuing to interact with the agent in Discord.
For my setup, Claude Code is the part that actually works in the codebase. You can inspect the repository, edit files, run commands, run tests, and work with the rest of the development environment. It would give you access to a local clone of the project instead of expecting you to do all the coding through a GitHub integration. This gives the agent a proper development cycle in which it can inspect the code, make a change, run the application or test suite, analyze the failure, and continue iterating until the task is completed.
I would still connect GitHub as a separate tool because it gives the agent access to the broader context around the codebase. GitHub has an official MCP server that can read repositories and code, work on issues and pull requests, and interact with workflows. This means that you can eventually send a message in Discord like: “Look at issue #142, find the cause of the bug, fix it, run the tests, and create a pull request.” The coding agent operates directly on the repository, while the GitHub integration helps you understand and manage everything related to that work.
Don’t forget the MD files.
Markdown files tell the agent how your project works
The next important step is to add Markdown files that tell the agent how your project works. This is what prevents you from tackling all the tasks, since you have never seen your codebase before. OpenClaw provides each agent with a dedicated workspace and automatically uses workspace files such as AGENTS.md, SOUL.md, TOOLS.md, IDENTITY.md, and USER.md as part of its context when present. It’s not necessary to fill every file with pages of instructions, but a few well-written files make a big difference.
I would start with AGENTS.md. This should explain what the agent is responsible for and how it is expected to function. I would also create a project specific file that explains the application itself. You could call it PROJECT.md and include the architecture, important directories, current priorities, known technical debt, and any areas the agent should avoid touching. OpenClaw does not automatically inject every arbitrary Markdown file into context, so it would explicitly reference PROJECT.md from AGENTS.md and instruct the agent to read it before doing substantial work.
If you are using Claude Code, I would also add a CLAUDE.md file to the repository. Claude Code uses CLAUDE.md files to store persistent project instructions, including coding standards, workflows, architectural notes, and must-know commands. This is separate from OpenClaw’s own workspace files.
You no longer need editors
Well, that’s a bit of a stretch, but for most use cases, you don’t need an editor anymore unless you really want to code things by hand. Basically, you can have your agent do things for you, and I think this is the direction publishers are moving in too. For example, VS Code now has an agent mode that doesn’t look like an editor at all. It just looks like a chat agent interface. Cursor has something similar and I’m sure this type of setup will become much more common very soon.
