Claude Tag and the context lock-in problem
On June 23, 2026, Anthropic shipped Claude Tag: a Slack integration where Claude joins as a teammate with access to the channels and tools you choose. Tag Claude into a thread, hand off a task, get back to what you were doing. The launch announcement is at 14M views.
A day later, Ashwin Gopinath wrote what we think is the right read on it:
Claude Tag is a Trojan horse. Not because Anthropic is doing anything evil. Because the incentives are obvious. ... That is not model lock-in. That is context lock-in. You are now renting your company back from them.
His whole thread is worth a read. We've been building Nairi for the same hypothesis Ashwin lays out: the part of an AI teammate worth owning is the context layer, not the model. Here's why we think that argument lands, and what the alternative architecture actually looks like.
Why context lock-in is the real risk
Models move. Cheap last year, free next year, replaced the year after. Switching from Claude to GPT-5 to Gemini to whatever ships next quarter is, for most teams, a config change. The intelligence is rentable.
What isn't rentable is the slow-built memory of how your company actually works.
The Slack scar tissue: the threads where someone tried something in Q2 and it didn't work. The exception paths: "we promised this customer net-90 even though our default is net-30." The implicit owners: who actually approves this kind of decision, who gets copied on this kind of email. The unfinished threads. The five-step workflow that lives in three people's heads. The "we tried that, here's why it failed" knowledge.
That memory is what turns an AI teammate from a generic assistant into something useful. It is also the part that, once it lives inside one vendor's product, is essentially impossible to move.
If Claude Tag becomes where that memory accumulates (channels, threads, tool calls, decisions, "ask Claude what we did last time"), then a year from now, when there is a model 30% cheaper or twice as good, you can't really move. You don't own the part of the system that compounded.
That's the argument. It's structural, not moral. Anthropic isn't doing anything wrong here. The incentives just point in a direction that's expensive for the customer over time.
What owning your context actually requires
It's tempting to read Ashwin's thread as a complaint and move on. The more interesting question, we think, is: what does it actually take to own your context?
A few things, by our reckoning.
Model portability. The agent's context can be reattached to a different model with the same fidelity. Today Claude, tomorrow GPT-5, next month an open-source model. Same memory of your org, different intelligence powering it.
Harness portability. The "harness" is the runtime that decides how the model talks to your tools. Claude Code, Codex, OpenCode, and Cursor each do this differently, and the choice changes the agent's behavior in ways that matter. Owning the context means you can swap which harness you run without re-onboarding the agent into your world.
Workspace-scoped memory. The org-level memory of how things work, rules, skills, MCP connections, vault secrets, channel permissions, lives in a workspace you control. Not a tenant inside the model vendor's account.
Inspectable and portable. You can read what the agent knows about you and your org. You can export it. You can take it elsewhere if you decide to.
No bundling of intelligence and workflow surface. The vendor selling you the model is not the same vendor running your Slack agent. If one of them slips, you can replace one without rebuilding the other.
Most "AI agent in Slack" products fail one or more of those tests by design, usually because the vendor that built the model also built the agent, and the bundling is the business model.
What this looks like with Nairi
We've been building Nairi against exactly that set of constraints. The short version.
Shared agents your team can use in Slack or Discord. Install the app, pick channels, mention an agent. Two minutes, no infra to deploy.
You pick the harness per agent. Claude Code, Codex, OpenCode, or Cursor. Bring the provider key you want. Want to swap from Claude to GPT or a local model? Change one setting on the agent. Your rules, skills, MCP connections, and vault come with it.
Workspace-scoped rules, skills, MCPs, and vault. Set once at the org level. Every agent inherits. None of it sits inside the model vendor's tenant. If you decide to leave Nairi, you export your config and you keep your context.
Self-hostable. The nairid daemon is open-source Go. The full agent loop, receiving Slack and Discord messages, running tools, posting back, runs in a binary you can host on your own hardware if you need to.
In Ashwin's language: we sell you the workflow surface, but the intelligence is yours to rent from whoever's best this month. We genuinely don't care which model you point your agents at. The product is the layer above the model.
Where Claude Tag wins
We try to write honest comparison content, so here's the part where we'd actually point you at Claude Tag instead of us.
Frictionless install if your team is already on Claude. If every engineer has a Claude seat and you're committed to Anthropic's roadmap, Claude Tag is one click away. No new product to onboard, no new vendor relationship.
First-party model and harness. No compatibility risk on Claude version bumps. Whatever Anthropic ships, you get on day one. Today that means Opus 4.8.
Ambient mode out of the box. Claude Tag can chime in unprompted, flag things from across the org, and follow up on forgotten threads. We have scheduled jobs and event triggers, but free-form passive channel observation isn't something we ship today.
Bundled pricing. No separate line item. It rolls into your existing Claude Team or Enterprise plan.
If you don't care about model portability or harness choice, and you're betting Anthropic stays the best forever, Claude Tag is the simpler choice. We won't try to talk you out of it.
The trade-off, going back to Ashwin's framing: the more value you push through it, the harder it gets to leave.
The architectural choice
The question for any team picking an AI teammate in 2026 isn't "which model is best today." It's "where is my company's operating memory going to live for the next five years?"
Rent the intelligence from whoever's best. Own the context.
That's the whole pitch, and Nairi is the platform we built to make it concrete. If you want to see what owning your context looks like in practice, try Nairi free or book a chat.
FAQ
What is Claude Tag?
Claude Tag is Anthropic's new Slack teammate, launched as a research preview on June 23, 2026. You tag @Claude in a Slack channel, Claude joins as a teammate, gets access to the channels and tools you grant, and you delegate work to it the way you would to a coworker. It runs on Claude Opus 4.8 on Anthropic's cloud, is gated to Claude Team and Enterprise plans, and replaces Anthropic's existing Claude in Slack app, which sunsets on August 3, 2026.
What does "context lock-in" mean?
The term, popularized by Ashwin Gopinath's thread on Claude Tag, refers to the structural dependency that forms when your company's operating memory builds up inside a single AI vendor's product. How your work actually flows, who owns what, exception paths, customer promises, "we tried that in Q2" knowledge. Models can be swapped. That context can't. Once enough of your work lives inside Claude Tag, you aren't really renting intelligence anymore. You're renting your company back from the vendor.
Is Nairi a Claude Tag alternative?
Yes. Nairi is a multi-vendor, workspace-scoped agent platform for Slack and Discord. You pick the harness (Claude Code, Codex, OpenCode, or Cursor) per agent, and the context (rules, skills, MCP connections, vault, channel permissions, memory) lives in your workspace, not inside the model vendor.
Can I still use Claude through Nairi?
Yes. Most Nairi customers do. Claude Code is the default harness for many agent types, and Nairi uses your provider key to bill Claude usage directly to you. The difference is that you can swap to a different model later without rebuilding the agent's context.
How is Claude Tag different from Claude Code for Slack?
They're related but distinct. Claude Code for Slack is Anthropic's developer-focused integration that opens PRs from a Slack thread under a per-user session. Claude Tag is broader: a persistent, multiplayer teammate that the whole channel shares, designed for any kind of work rather than just code. Anthropic frames Claude Tag as "the beginning of an evolution of Claude Code." If you want a structured side-by-side, our Nairi vs Claude Tag page goes through the dimensions row by row.
What if I'm already on Claude Tag?
A fair experiment: pick the highest-value workflow you're already running through Claude Tag, replicate it on Nairi with whichever harness fits best, and see whether the multi-vendor architecture actually matters for that workflow. If it doesn't, you've lost an afternoon. If it does, you've found out before the lock-in compounded.
Where can I read Ashwin's original thread?
Ashwin Gopinath on X, June 24, 2026. The "rent the intelligence, own the context" framing is his.