Open-source alternatives to Claude Tag, by the layer each covers

Three open options, each covering a different layer

Kortix is the open-source AI Operating System, and it fills the layer under the chat surface: your agents, their skills, your company memory, and every connector in one git repo you own, with the agents working on real cloud computers. A project is a git repo with a kortix.yaml at its root. Agents, skills, memory, tools, triggers and connectors are files inside it. One isolated sandbox per session: each session has its own isolated machine and branch, and its work lands through a change request a person reviews.

The open-source alternative to Claude Tag that exists today is open-tag, and it covers the chat surface instead. It is a self-hosted workspace rather than a Slack app: a React web app, a server, and a daemon on machines you own. People and agents share channels, threads, DMs and a task board, and you reach it at an address you control. The project's own README calls it "the open-source alternative to Claude Tag".

Mention an agent in a channel and the daemon starts a CLI runtime on a machine you have registered. The README lists Claude Code, Codex, Copilot, OpenCode, Kimi Code, Pi and Cursor as supported, with Hermes and Reasonix marked experimental. Each agent keeps a persistent workspace and its own MEMORY.md. The project describes itself as early-stage software: authentication and deployment suit self-hosted evaluation, it says, while production hardening, third-party OAuth integrations, web push and large multi-host deployments remain active work.

Claude Tag sits at the other end, and it is closed. Anthropic's support page describes a product that works in Slack today, on Team and Enterprise plans in beta, through channel mentions, direct messages and the assistant panel. A channel mention runs under the organization's identity, with the tools an admin provisioned, and is billed to the organization. Memory is kept per channel and per workspace. An owner can review, edit or delete it.

The three do not answer the same question. Claude Tag and open-tag both answer where the conversation happens. Kortix answers what the agent runs on: which repo holds its skills and memory, which machine it works on, which connectors it may call, and whose model replies. That is the layer a team is really buying when it asks for an open Claude Tag. The bot in the thread is the visible part. The machine, the repo, the connectors and the model underneath decide what the bot can do.

What open source buys a team in this position

Open source. Read it, fork it, audit it. That sentence matters most at the runtime layer. The runtime is where an agent gets its reach: which connectors it may call, which secrets it may read, which files it may change.

open-tag publishes its code, so you can trace what the daemon does with a mention and where an agent's workspace ends. Kortix's code is public too, at Kortix on GitHub, so the permission model is code you can open instead of a paragraph you have to trust.

The second thing you buy is the exit. A workspace you host keeps its messages, task board and per-agent memory in a database and a directory you can copy. A runtime you host keeps skills and memory as files in your repo, so changing hosts is a clone and a push.

What stays your problem when you self-host

Self-hosting changes who carries the operations, and the operations are the honest part of the comparison. Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free. On your own box, the list is concrete.

You need a Linux server with a public IP, DNS records for the domain and the API subdomain, a wildcard record for sandbox previews, and ports 80 and 443 open. A Platinum API key is required, and without it sessions do not start. Each project also carries its own model provider key, because a self-hosted instance has no Kortix-managed models. Set an email connection before you invite anyone; until then a new instance confirms every sign-up without proving the address. Back up .env, the database volume and stored files, since .env holds the key that encrypts every stored secret. kortix self-host doctor reports which of those is still wrong.

An open-tag install has its own list: Node 20 or later, Docker, at least one runtime CLI on the PATH, two generated secrets in .env, a database, and at least one project root the daemon may bind. Its README points a production host at Docker Compose behind HTTPS with systemd and backups. It also says plainly that production hardening is still in progress.

Models, connectors, and the choice in front of you

Model choice follows the same split. Kortix puts it in the project: Any model provider with your own keys. Or the ChatGPT plan you already pay for. Or sign in with your OpenCode Console account for OpenCode Zen and Go. open-tag inherits whatever each CLI runtime is configured with. Claude Tag runs Claude models. Reach into other systems comes from connectors on all three, and on Kortix each one carries a policy of allow, ask or block.

The choice comes down to which layer is missing. Kortix fills the runtime layer: agents, skills and memory as files in a repo you own, a cloud computer per session, and Slack or Teams as one way in. Neither of the other two covers it.

A team that wants the conversation surface on its own server, and already runs coding agents on machines it controls, gets channels, tasks and a daemon from open-tag. It keeps supplying the machines and their credentials.

If Claude in Slack with the least setup is the goal, and your organization is already on Team or Enterprise, Claude Tag is the shortest path. The trade is that the runtime, the memory and the model stay on Anthropic's side.