Claude Tag in Slack: August 3 retires the old app — migrate or get switched

August 2, 202615 minClaude, Anthropic, Slack, AI agents, enterprise SaaS, Latam

Short answer (60 seconds): Anthropic launched Claude Tag on June 23, 2026 — what several analysts call the first shared AI teammate in Slack (per-channel, not per-user) — with persistent memory, channel-scoped identity and tool access via MCP (anthropic.com/news/introducing-claude-tag). It runs on Opus 4.8 and replaces the legacy "Claude in Slack" app, which retires on August 3, 2026 (computerworld.com). If you do not opt in manually before then, Claude in Slack will be switched over to the new Claude Tag experience on August 3, 2026. There is an introductory credit of USD 25K per Enterprise org or USD 2.5K per Team org (10+ seats), expiring September 1, 2026 (support.claude.com). For SaaS founders selling "agents in Slack" or "AI coworker", this redefines the UX pattern your customer now expects. This week is the time to audit your position.

This is the post I wish I had on a Friday afternoon when an enterprise client asked "should we keep selling our Slack integration or let them migrate to Claude Tag?". The short answer: it depends on whether your SaaS is data/context or execution/decision, the same framework I applied to ChatGPT Work (prior post). But the difference is that Claude Tag is not an agent layer on top of your product — it is a competitor that ships with Opus 4.8 + MCP + native Slack. The decision is this week, not next month.

Disclosure: I reviewed the official Anthropic announcement, the DataCamp breakdown, the beri.net enterprise guide, and the digitalapplied.com team collaboration analysis. I have not installed Claude Tag in my own workspace yet — the unit economics numbers are modeled on Opus 4.8 public pricing and the published credits. Where I cite a figure without an exact source, I flag it inline.

What happened this week (and why it matters)

Anthropic launched Claude Tag on June 23, 2026 (anthropic.com) — what several analysts call the first "AI teammate" that lives inside a Slack channel rather than as a 1-on-1 chatbot. The key architectural decision: a single @Claude per channel that anyone can tag, that maintains shared memory across humans and tools, and that behaves proactively. If you have the legacy "Claude in Slack" app installed, August 3, 2026 is when it retires (computerworld.com). You had 30 days to migrate; Claude in Slack will be switched over to the new Claude Tag experience on August 3, 2026.

Note as of August 2: Anthropic launched Opus 5 on July 24, 2026 (9 days before this post) at the same price as Opus 4.8 ($5/$25 per million input/output). As of this writing, Anthropic has not officially upgraded Claude Tag to Opus 5 — the experience still runs on Opus 4.8. This does NOT change the unit economics in this analysis: the token figures, launch credits (USD 25K / USD 2.5K) and per-channel math still hold because Opus 5 keeps the same public pricing. If Anthropic ships the upgrade in the next few weeks, customers will not see a material change in their bill.

Four launch facts that matter for the SaaS founder:

1. It is per-channel, not per-user. The legacy app was a chat with Claude per person. Claude Tag is a channel participant with its own identity, memory and permissions. Anthropic says internally "65% of our product team's code is created by our internal version of Claude Tag" (anthropic.com). The analogy: it is not "asking a chatbot a question", it is "delegating to a teammate".

2. It has persistent memory and proactive behavior. Claude Tag accumulates context about the channel's work, can learn from other channels it has permission to read, and — if you enable "ambient" mode — flags things it considers relevant even when nobody tagged it (anthropic.com). The key word is proactivity — the legacy app waited for you to ask; this one notices on its own and offers.

3. Strict scoping per channel. The admin defines which tools, data sources and codebases each Claude can see in each channel. A "sales" Claude does not share memory or access with an "engineering" Claude (anthropic.com, Getting started section). This matters for compliance: you do not need to re-architect for multi-tenancy, the product ships with scoped identity out of the box.

4. Aggressive launch credit. USD 25,000 per Enterprise organization, USD 2,500 per Team with 10+ seats (support.claude.com, computerworld.com). Expires September 1, 2026. Your free-trial window is 30 days, not 60 — read that twice before moving on.

Claude in Slack (legacy) vs Claude Tag (new) — comparison table

Here is the difference that matters when evaluating whether to migrate or compete:

Row-by-row reading:

• Per-channel identity is the paradigm shift. It turns Claude from a personal tool into team infrastructure. If your SaaS sells "AI copilot per user", this row alone forces you to rethink pricing.
• Persistent memory redefines the value proposition: the customer does not train a new bot per channel, they adopt a teammate that already knows the work.
• Ambient mode + async erases the line between "automation" and "collaboration". Today you schedule a Zap; tomorrow you tag Claude and it remembers to do the recurring work.
• MCP means any tool your SaaS exposes via an MCP server can be used by Claude Tag without custom integration. If your SaaS does not yet have an MCP server, this week is the week to ship one.
• Per-channel spend limit gives you granular financial control. For the first time you can cap how much Claude "spends" in a specific channel without cutting off access for the rest of the org.
• Pricing is still not published per-seat. What you have is your Enterprise/Team subscription + tokens consumed at Opus 4.8 pricing. The credit covers your trial; afterwards, budgeting matters.

What "per-channel Claude" means for your unit economics

The shift from per-user to per-channel is not cosmetic — it changes the math. Here is the comparison with concrete numbers.

Base case: B2B SaaS with 50 active customers. Each customer has 10-20 Slack channels where their team works (incidents, features, on-call, customer success, sales). 15 channels per workspace on average.

Legacy model (Claude in Slack, per-user):

• If your customer pays USD 30/seat/month and has 50 people using Claude in Slack, Anthropic's bill runs per active seat.
• In a team of 50, if 30 actively use Claude in Slack, the customer pays ~USD 900/month to Anthropic at per-seat pricing (estimated; exact per-seat pricing is not published, this is the typical category pattern).
• The inference cost per workspace scales with the number of people who use it.

New model (Claude Tag, per-channel):

• The bill is token-based on Opus 4.8 ($5/$25 per million input/output).
• Per active channel with moderate usage (estimated: 50 messages to Claude/month, ~500 tokens input + 300 output per message), the cost is:
  • Input: 50 × 500 = 25,000 tokens → $0.125
  • Output: 50 × 300 = 15,000 tokens → $0.375
  • Total: ~$0.50/channel/month at Opus 4.8 pricing without cache
• With 50% cache hit rate on system prompt: ~$0.30/channel/month
• For 15 channels per workspace: USD 4.50 to 7.50/workspace/month in tokens

Compared to the per-user model:

Calculation disclaimer: figures assume Opus 4.8 at public pricing, system prompt of ~600 tokens, average message of 500 tokens input + 300 output, 50 messages/channel/month. Your real number depends on the system prompt, the type of tasks your customer tags Claude for, and how much they use ambient mode. Before making decisions, run your own pilot with your real workload.

What this implies for your SaaS:

1. If you sold "AI copilot per seat", the customer's pricing reference just changed. Your USD 30/seat differential is now measured against a Claude that costs fractions of a cent per channel. Your moat has to live in something other than model capability.
2. If you sold to enterprise customers with CRMs/ERPs/internal codebases, you now compete with Claude Tag + MCP. Your value has to be data, workflow or compliance — not integration.
3. The launch credit is an aggressive free-trial window. USD 25K Enterprise at Opus pricing buys a huge volume of tasks. If your customer tries it, they will compare your pricing against "this costs me nothing for X months". Model how you defend that.

What to audit in your SaaS this week

Operational checklist to run between Monday and Friday. If your SaaS touches Slack, AI, or "AI teammates", this week is yours.

1. Do you compete with "AI teammate" or "AI coworker"? If your value prop says "agent", "teammate", "AI coworker" or "AI employee", Claude Tag redefines the pattern. The customer no longer compares you to other SaaS in your category — they compare you to a native @Claude in their Slack. Three questions: (a) what does your agent do that Claude Tag with MCP cannot; (b) how much of your pricing is the agent layer vs proprietary data/workflow; (c) what happens if your customer adopts Claude Tag and keeps you only for the data.

2. Does your SaaS integrate with Slack? If you have a Slack integration (notifications, slash commands, conversational bot, modal, app home), you have three paths: (a) expose your product via an MCP server so Claude Tag can use it as a tool — the fastest path; (b) compete on superior UX and make your bot better than @Claude; (c) hide and wait for the pattern to settle. Hiding is the worst option: by the time it settles, your customer has already decided.

3. Should you build on MCP or on the Anthropic SDK? MCP (Model Context Protocol) is the open standard Anthropic launched for tools to connect to models. Claude Tag uses it out of the box. If your product exposes an MCP server, Claude Tag discovers and uses it without writing a line of integration code. It is 10x less friction than building a Slack app from scratch. If you do not yet have an MCP server, ship a minimal one before the end of the month.

4. Do you have visibility into per-channel spend? Most SaaS does not have this. What you need: a dashboard that shows tokens consumed by (channel, day, customer) over the last 30 days. If you do not have it, two hours with the API log and a warehouse query get it running. This lets you benchmark against Claude Tag when your customer asks "what if I just use Claude Tag directly instead of your product?".

5. Is your pricing per-user or per-channel? If your pricing is per-user and your competitor (Claude Tag) is per-channel, there is an incentive mismatch that will eat your margin. Either move your pricing to a model that reflects value per channel/workspace, or you will lose in enterprise deals where the CFO asks about the cost difference.

When NOT to use Claude Tag

Claude Tag is competitive but not a universal replacement. Three cases where it makes sense to keep the migrated legacy app, or not adopt it at all:

1. Channels with heavy compliance (HIPAA, FINRA, gov/defense). Claude Tag inherits Anthropic's policies, but your customer may have data residency requirements (LATAM banks, healthcare, government) that a US-hosted model does not meet. If your SaaS competes in these verticals, you remain the compliance moat. Verify with your legal team before promoting Claude Tag as a replacement for your product.

2. Financial data / quantitative modeling where 1 error in 1000 matters. Opus 4.8 is still an LLM. For critical financial decisions (portfolio rebalancing, liquidations, tax calculations, legal contracts) the cost of a single error is measured in seven figures. If your product competes in this range, your moat is the deterministic model + audit trail, not the AI. Claude Tag can be a support layer (summarize, search, draft) but not the decision system.

3. Workloads with system prompts above 50K tokens. Claude Tag runs on Opus 4.8, which has a 1M context window. But past ~50K tokens of system prompt repeated per channel, input cost per message becomes material without a high cache hit rate. If your product relies on large contexts (RAG with 100K+ chunks, entire monorepo codebase), Claude Tag is not optimal as the primary host — use it as the orchestration layer, not the retrieval backend.

Three operational practices that apply wherever you are

Regardless of whether you adopt Claude Tag or compete against it, three practices reduce operational risk:

1. Ship a minimal MCP server before the end of the month. This is not a 6-month project. Take your product's most-used tool/API, expose it as an MCP server with 3-5 tools, and publish it in your docs. When a customer asks "how does this integrate with Claude Tag?", the answer becomes "there is an official MCP server, install it and you are done". It is 10x cheaper than maintaining a custom integration and opens the door to your product being one more tool in the customer's ecosystem rather than a locked-out competitor.

2. Separate "what happens to my data" from "what happens to my agent". If your SaaS is data/context, let Claude Tag query your backend with a per-API-call charge. If your SaaS is execution/decision, do not expose your decisions via MCP — keep them, and charge the customer for the decision, not the access. It is the same framework I applied to ChatGPT Work (reference): integrate or compete depends on where your moat lives.

3. Measure cost per channel, not just per seat. The "bill per active seat" pattern hides how efficient your product is at the level of real usage. With Claude Tag as a benchmark, your customer can now ask "how much does Claude Tag cost for my volume vs your product for my volume?". If you cannot answer with numbers, you lost the deal.

Conclusion

Claude Tag is not another product in the category — it is a pattern shift. We moved from "AI per user" to "AI per channel" with shared memory, proactive behavior and tools via MCP. The legacy app retires August 3; the launch credit expires September 1. If your SaaS touches Slack, AI, or sells "AI teammates", this week is when you decide.

Three operational recommendations to close the month:

1. If you sell "AI teammate" or "AI coworker": define this week whether your moat is data/workflow or execution/decision. If it is the latter, ship an MCP server and let Claude Tag be the agent layer on top of your product. If it is the former, do not integrate — compete on UX.
2. If your SaaS integrates with Slack: publish a minimal MCP server before the end of the month. It is the cheapest integration you can ship to stay inside the pattern your customer is already evaluating.
3. Run the unit economics math with the credits: USD 25K Enterprise or USD 2.5K Team gives you an aggressive free-trial window. Model how many channels/month of your workload fit inside that credit and what the next month costs. If the answer makes you uncomfortable, you need to reposition pricing before the customer does it for you.

If your SaaS startup is evaluating how to position against Claude Tag or wants to validate the integrate/compete decision with your real numbers, book a free 30-minute call — we usually unblock the pricing/positioning decision in one session.

---

Read also:

• OpenAI ChatGPT Work vs your SaaS: how to position before it becomes commodity — the same integrate/hurt framework, this time against ChatGPT Work.
• MCP explained for SaaS founders — what the Model Context Protocol is and why it matters for Claude Tag.
• Claude Sonnet 5 at $2/$10: review your COGS this week — Claude Tag runs on Opus 4.8; the model's unit economics still matter.
• Claude Opus 5: what changes versus Opus 4.8 — the tier above the model, in case your workload benefits.
• Back to the blog — all articles.

Frequently asked questions

What is Claude Tag in one sentence?

It is a shared AI teammate from Anthropic that lives inside Slack channels: a single @Claude per channel with persistent memory, channel-scoped identity, proactive behavior and tool access via MCP. It runs on Opus 4.8. It replaces the legacy "Claude in Slack" app and is in beta for Claude Enterprise and Team customers.

When does "Claude in Slack" retire and what if I do not migrate?

August 3, 2026. You had a 30-day window from June 23 to opt in manually. If you did not, Claude in Slack will be switched over to the new Claude Tag experience on August 3, 2026 and the legacy app stops working. The channel-level upgrade is transparent (same @Claude handle), but the legacy app's configuration does not carry over — so identity, connected tools and spend limits all need to be re-set.

How much does Claude Tag cost?

The feature itself has no published per-seat or per-feature fee. The real cost is token-based: the admin sets a monthly spend limit at the organization and channel level, and Claude Tag runs on Opus 4.8 with the same per-token billing as the API. To get started, Anthropic is issuing an introductory credit of USD 25K per Enterprise org or USD 2.5K per Team org (10+ seats), expiring September 1, 2026. After the credit, you pay tokens at Opus 4.8 pricing ($5/$25 per million input/output).

Is Claude Tag per-user or per-channel?

Per-channel. That is the central architectural shift. The legacy app was a 1-on-1 chat with Claude per user; Claude Tag is a channel participant anyone can tag, that sees what other humans requested, keeps shared memory, and can also chat in private DMs with the tools you configure for it. The closest analogy is "an actual teammate" rather than a chatbot.

How do I connect Claude Tag to my internal tools?

Via MCP (Model Context Protocol). The admin defines which MCP servers and which scopes each Claude has in each channel. For example, a "sales" Claude can read your CRM but not your codebase; an "engineering" Claude can read GitHub but not Salesforce. Scoping is strict: memory, tools and data sources are isolated by channel, so they do not leak between identities.

How does Claude Tag impact a SaaS competing with "AI teammates"?

It redefines the category. If your product sells "AI agent per user" or "AI coworker per seat", Claude Tag shifts the pattern: the customer now expects their agent to be shared in the channel, with persistent memory and proactive behavior. Your moat has to come from something other than the orchestration layer — proprietary data, native workflow, or strict compliance. If your product lives in Slack, you need to decide quickly whether to integrate Claude Tag (via MCP or API) or compete on superior UX. Waiting is not an option.