Claude Tag vs Kortix: open source is what you keep

One request, run in two workspaces

The request is the same in both tools: last quarter's renewal numbers out of the CRM, a draft of the customer note, and the deck on a branch. At Acme, member 4 types it into #revenue-ops on a Tuesday morning.

On Claude Tag, most of the setup happened before that message. A Primary Owner provisioned Claude's identity for the channel and connected the CRM, because only that role can. Member 4 tags @Claude. Claude splits the job into stages, calls the tools the channel was granted, and answers in the thread. The next morning a different member picks the thread up, and Claude still has the context. One channel holds one Claude, and channel memory is kept rather than discarded. Owners can read, edit and delete what it remembers. The Activity page lists the task along with the network calls made under the organization's identity.

On Kortix, member 4 sends /kortix login once per workspace and then mentions the bot. One Slack thread is one session. The first message starts it, later replies continue it, and the session keeps its state when the sandbox stops and resumes. The agent acts with its own grant, separate from member 4's project role.

Two replies land in two threads by lunchtime. The difference shows on Thursday, when a reviewer asks what the agent can still reach, and again on Friday, when someone asks where the memory sits.

What lands, and who can hold it

A Kortix session does more than answer in a thread. When it files a change, a card arrives with Ship it, Reject and Ask for changes. The same item waits on the Review page in Kortix. A connector call that policy holds posts its own card: the action, the parameters the connector will receive, and the agent's note on the effect, marked as not verified.

That review path is the one the product itself ships on. A change request merged on 23 October 2026 fixes how a scoped token resolves its own project before authorization runs. It reached the trunk with tests, a self-review and a rollback line, like every other change (the merged change request).

Claude Tag finishes in the thread, and its controls sit in the admin console. Member access has three modes. Credentials and repositories are set at organization, workspace or private channel level. Spend carries an organization cap, per-channel limits and alerts at 75 and 95 percent. Commits made through the GitHub connection show the Claude GitHub App as author and link back to the thread that asked for them.

What you still own in year two

The reason to compare these two at all is what a company keeps once the pilot ends. The answer is mostly about where the memory and the runtime sit.

Claude Tag keeps memory per channel and per workspace, retains it, and lets owners view, edit and delete it. That is real control, and all of it is exercised inside Anthropic's platform, through Anthropic's admin screens. The runtime and the model stay there as well. For a team whose agent answers questions in a few channels, the trade is fair and the setup is quick.

Kortix files the same things as code. Your agents, their skills, your company memory, and every connector live in one git repo you own. A memory that turns out to be wrong is a diff someone can review and revert. Any model provider with your own keys. Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free.

The renewal history, the support patterns and the runbook are the assets here.

When Claude Tag is the right pick

Claude Tag wins a real case, and it wins it on effort. The team already pays for Claude Team or Enterprise, works in Slack all day, and wants a shared teammate without standing anything up. The security review can lean on the vendor relationship the company already has. Administrators get member access, per-channel reach and spend caps without a runbook.

If that describes you, Claude Tag is the better pick. It is in beta, it works in Slack today, and Anthropic says it plans to widen where it runs. You want Anthropic's models specifically, and nobody has asked where the compute sits.

The choice flips when the agent will hold company memory across years, act in the systems a reviewer cares about, or outlive the current roadmap. Then the runtime, the model choice and the review path decide it.