Introducing openTag

Why we built it

Over the past year, coding agents like Claude Code and Codex quietly became the most capable “workers” on our team. They can read a codebase, query a database, write a report, open a pull request, and check their own work. But they live in a terminal on one person’s laptop. To use them, you have to be at that laptop, in that terminal, babysitting that session.

Meanwhile, the work itself arrives in Slack: a bug report in #support, a metrics question from a PM, a “can someone look into this?” in a team channel. We wanted to hand that work to an agent right where it shows up, and get the answer back in the same thread.

When we looked at the AI workers already on the market, they didn’t fit how we work. Most are closed-source and hosted only in someone else’s cloud. They come with their own agent runtime and their own model contract, so you pay a second time for intelligence you already have a subscription for. And your files, credentials and conversations end up on infrastructure you don’t control.

We wanted the opposite: something lightweight and open, that runs on our own machines, uses our favorite agent harness and models, and that we can bend to our own workflows. So we built openTag.

How it works

openTag has three pieces:

• Your team chat. You talk to openTag workers in Slack (Lark is supported too), in a channel or a DM.
• The openTag server. It receives messages from Slack, keeps track of workers, tasks and sessions, and routes each task to the right computer. You can use our hosted server or run your own; it’s fully open source.
• A small daemon on your computer. You connect a Mac or Linux machine by pasting one command into a terminal. The daemon keeps a connection to the server, and when a task arrives it starts a session with your local agent harness, passes in the thread’s context, and posts the result back to Slack.

From your point of view, that’s it. You @mention a worker, like “@openTag how many recurring users did we have last week?”, and it picks up the task, works in the background with the context of the channel it’s in, and replies in the thread when it’s done or when it needs more input from you. Follow-up messages in the same thread continue the same session, so you can iterate with a worker the way you would with a colleague.

How openTag is different

1. It works with your favorite agent harness, and your subscription

openTag is not yet another agent harness. We didn’t build a new agent loop, a new tool system, or a new model wrapper. Instead, openTag puts the harness you already use to work and connects it to Slack. Today that means Claude Code, Codex, and Pi, and we’re adding more.

This has two big consequences.

First, openTag is model-agnostic. Whatever models your harness can run, your openTag worker can use. Want one worker on GPT and another on Claude? Pick the harness and model per worker. When a better model ships, you get it as soon as your harness supports it, with no waiting for us to integrate it.

Second, you can use the subscription you already pay for. Because openTag drives your own harness installation, signed in with your own account, the work runs on your existing plan. OpenAI lets you use your ChatGPT subscription with third-party apps, so you can run an AI worker on the tokens already included in your ChatGPT plan, instead of paying for a separate per-seat AI product or a new API bill.

2. It runs on your computer, and your data is yours

By default, every openTag worker runs on a computer you own: your laptop, a Mac mini under a desk, or a Linux box in your office. That means a worker can use the same local files, repositories, CLIs and tools you do, without uploading your codebase or your credentials to someone else’s cloud.

It also means your data stays with you. Agent sessions and working files live on your machine. Workers can share knowledge through a git-backed, file-based memory: plain markdown files in a git repository that you own, which you can read, edit, review and roll back like any other code.

And you stay in control of what the agent does. On a local computer, Claude Code and Codex workers ask before risky actions: openTag sends a private approval card to the person who started the task, with Approve and Deny buttons, and the agent only continues once you decide. Only the person who started the task can approve its actions.

3. It’s highly customizable

Every team works differently, so every openTag worker is yours to shape. Each worker has its own instructions, model, skills and MCP servers.

• An engineer can be hooked up to GitHub, Linear and Sentry.
• A saleslead can connect to your CRM and calendar.
• A designer can connect to Figma.

openTag connects to any standard Agent Skill and MCP server, so the whole ecosystem of existing integrations works out of the box. And you don’t have to wire them up by hand: you can simply ask your worker in Slack, like “add the Linear MCP” or “install a skill for reviewing PRs”, and it configures itself. The same goes for its instructions and model.

Because openTag is open source under the Apache 2.0 license, you can go further: read the code, change it, and self-host the entire stack. The whole stack is self-hostable, including the server that connects Slack to your workers.

Prefer not to keep a computer running? Use our cloud

Running workers on your own computer is the default, but it does mean that computer has to stay on and online while your workers are working. If you’d rather not manage that, openTag Cloud runs your workers in a cloud sandbox, with model tokens included.

There’s nothing to install and no machine to keep awake. You create a worker, connect Slack, and start delegating. Pricing is simple: you pay for the tokens your workers use.

You can mix and match, too: keep the workers that need your local files and tools on your own computer, and put always-on workers in the cloud.