Claude Slack Integration: Tag or Custom Bot?

There are now two serious ways to put Claude to work inside Slack. You can turn on Claude Tag, Anthropic’s own Slack product, or you can build a custom bot on the Claude API. They suit different businesses. This post sets out what each one does, where each one stops, and how to tell which one you need.

We build Claude integrations through our AI integration service, so weigh that. Every fact below about Anthropic and Slack was checked against their own documentation in October 2026.

Short answer: If your company is on a Claude Team or Enterprise plan, has no Zero Data Retention or HIPAA requirement, and mostly wants a shared assistant in channels, start with Claude Tag. Build a custom Claude API bot when you’re on an individual Claude plan, when you have compliance needs Claude Tag excludes, or when you need one fixed workflow that behaves the same way every time. Either way, the hard part is deciding what the assistant should read, produce and refuse.

Claude and Slack in October 2026: The Facts

As of October 2026, from Anthropic and Slack documentation:

Sources: Claude Tag setup, Claude Tag spend limit, Migrate from the earlier Claude in Slack, Claude Tag security and data, Claude Tag for healthcare, Slack Events API, conversations.history, Socket Mode, Claude pricing, model training, data retention.

Native Slack App vs. Custom Claude API Bot, What’s the Real Difference

“The Claude app in Slack” now means two different things. Anthropic first announced Claude in Slack in October 2025 (source). That earlier app linked each person’s own claude.ai account. The current product is Claude Tag, which gives the whole team one Claude identity set up by an admin (source). Slack’s help centre says Claude Tag replaces the earlier app starting August 3, 2026 (source).

A custom bot is something else again. Your own code receives Slack events, calls the Claude API and posts the reply. You pick the model, write the instructions and decide what data it sees.

What the Native Claude App in Slack Can and Cannot Do

The earlier app was a personal assistant. Anthropic’s own comparison lists it with no memory and no standing work, with a new session spawned for each request (source). It also needed a paid Slack plan and a Slack admin’s approval (source).

Claude Tag closes most of those gaps. According to Anthropic’s docs:

• It keeps per-channel memory, plus workspace notes shared from public channels (source).
• It can run routines on a schedule and watch channels (source).
• An Owner sets it up once and decides what it can reach in each channel (source).
• Each Slack thread runs in its own ephemeral sandbox (source).
• Channel work bills to your organization’s usage balance under a spend limit you set, and members can use it in channels without holding a Claude seat (source).

What it can’t do is just as clear. It isn’t available on Free, Pro or Max plans (source). It isn’t available to organizations with Zero Data Retention, customer-managed encryption or the HIPAA configuration, because it keeps channel memory and transcripts (source). It isn’t covered by Anthropic’s BAA. By default it can search every public channel by keyword, including ones it hasn’t been added to, and no admin setting turns that search off (source).

What a Custom Claude API Integration Actually Gives You

A custom build trades convenience for control. You decide:

• Inputs: which channels the bot listens to, and which systems it may query.
• Outputs: the exact format of every reply, and where it gets posted.
• Behavior: what it refuses, and when it hands off to a person.

With tool use you can set strict: true so Claude’s tool calls match your schema exactly (source). That matters when the bot writes to another system.

One old claim needs correcting. A custom build does not keep processing on your own servers. Your code runs where you host it, but the model still runs on Anthropic’s side. What changes is the retention terms. API inputs and outputs aren’t used for training by default (source), and they’re deleted within 30 days, with exceptions such as files you upload, Zero Data Retention agreements and usage-policy enforcement (source).

What an Internal Business Assistant Should Actually Do in Slack

The useful question isn’t “what can AI do in Slack”. It’s “which of our workflows are slow, error-prone, or stuck in one person’s head”. Those are the ones worth handing to an assistant, whether that’s Claude Tag or your own bot.

Workflow Automation Examples That Move Real Work

These are common shapes, not results we’re claiming:

• Weekly status digests: on a schedule, pull open items from your project tool and post a short summary per client or project to a private channel. A person reviews it before it goes anywhere else.
• Proposal drafts from CRM data: when a deal reaches a set stage, gather the contact record and scope notes and draft an outline for the account manager to finish.
• HR and policy answers: staff ask in one channel and the bot answers from the handbook, citing the section it used.
• Internal onboarding: new hires ask where a process lives or who owns what, and the bot answers from your wiki rather than general knowledge.

A bot gives good answers when it’s scoped to one repeatable job. When it isn’t, it often has to admit it can’t see the information. Scoping is the real work.

What It Should Refuse to Do, Guardrails Are Not Optional

An assistant that answers everything with confidence is a liability. Your instructions should spell out which sources it may use, which actions it may take, and which topics it must pass to a person.

An HR policy bot should refuse to interpret employment law or advise on a dismissal. A proposal bot should flag work outside the services you’ve defined instead of guessing. These limits are what make people trust the tool enough to keep using it.

Guardrails also cover what the bot can see. Claude Tag’s keyword search reaches every public channel by default, and Anthropic’s healthcare guidance says to keep protected health information out of all public channels in the workspace for that reason (source). A custom bot only sees the channels and scopes you grant it.

How to Build a Claude API Slack Integration for Your Business

This isn’t a full coding tutorial. It covers the parts that decide whether the build works: the architecture, the Slack limits and the cost.

Architecture Basics: Slack Events API + Claude API + Your Data Sources

A custom bot has three layers:

1. Slack delivers events. With the Events API, Slack sends messages and mentions to a public URL on your server. Your app must answer with HTTP 2xx within three seconds, or Slack treats the delivery as failed and retries it three times (source). Because Claude can take longer than three seconds, acknowledge first and do the work afterwards. If your app sits behind a corporate firewall, Socket Mode uses a WebSocket instead of a public URL, but Socket Mode apps can’t be listed in the Slack Marketplace (source).
2. Your backend gathers context and calls Claude. It fetches what the question needs from your systems, builds the prompt and sends it to the Claude API. Slack’s Bolt framework handles much of the Slack side for you (source).
3. Your backend posts the answer. chat.postMessage needs the chat:write scope, and Slack allows about one message per second per channel (source). For longer answers, Slack offers streaming through chat.startStream, chat.appendStream and chat.stopStream (source).

Reading history is where Slack’s limits bite. For internal and Marketplace apps, conversations.history has Tier 3 rate limits. Since May 29, 2025, new apps distributed commercially outside the Marketplace are limited to 1 request per minute and 15 objects per call (source). A bot built for your own workspace is an internal app, so the stricter limit doesn’t apply. Still, fetching only what each question needs beats bulk reading.

The data-source layer is where most of the work lives: connecting your CRM, keeping documents current and keeping prompts focused.

Defining Inputs, Outputs, and Behavior Before Writing a Line of Code

Most teams skip this step. Before any code, write down what the bot receives, what it produces and what it refuses. A short spec answering those three questions gives the developer something to test against.

Inputs: Which channels does it listen to? What outside data does it fetch, and how fresh must it be? How much context does a typical question need?

Outputs: What format does the reply take? Who sees it? Does it reply in the thread, in the channel, or by direct message?

Behavior: What does it say when a question is out of scope? How does it show uncertainty instead of making up an answer? Who does it hand off to?

The same spec helps if you choose Claude Tag. You’d use it to set each channel’s connectors and instructions instead of writing code.

Hosting, Costs, and Maintenance, The Numbers Nobody Mentions

Claude API costs depend on tokens, so you can estimate them before you build. Anthropic lists Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, and Claude Haiku 4.5 at $1 and $5 (source).

Our own worked example, using example assumptions rather than client data:

• HR policy bot: 200 questions a month, each sending 6,000 input tokens and getting 400 output tokens back. That’s 1.2 million input tokens ($2.40) and 80,000 output tokens ($0.80), so about $3.20 a month on Sonnet 5.5.
• Weekday digest: 22 runs a month, each reading 20,000 tokens and writing 1,500. That’s 440,000 input tokens ($0.88) and 33,000 output tokens ($0.33), so about $1.21 a month.

On Haiku 4.5 both figures halve. If the same handbook goes into every prompt, prompt caching helps: a cache hit on Sonnet 5.5 costs $0.10 per million tokens (source), and the cache lasts 5 minutes by default, with a 1-hour option (source). Jobs that don’t need an instant answer can use the Batch API at 50% off.

Token spend is rarely the biggest number. Hosting, development and upkeep matter more, and they depend on your stack, so get them quoted against your spec. Expect maintenance too. Slack and Anthropic both change their APIs and models, and your data sources change as well.

Claude Tag costs work differently. There’s no bot to host. Usage bills to your organization’s balance, capped by the spend limit you set (source).

Who Should Build This vs. Who Should Wait

Not every business needs a custom bot, and some don’t need a Slack assistant at all. Saying so upfront saves everyone time.

Signs Your Business Is Ready

• You have at least one workflow that repeats: same inputs, same steps, same output. And it takes real staff time.
• Someone, in-house or an agency, can maintain the bot when Slack, Anthropic or your own tools change.
• Claude Tag doesn’t fit your setup. Maybe you’re on an individual Claude plan, or you need Zero Data Retention, customer-managed encryption or the HIPAA configuration (source).
• You need fixed behavior: the same model, prompt and data flow on every run, written into code you can test.

When the Native App (or No App) Is the Right Call

Claude Tag is the better first step if:

• You’re already on a Claude Team or Enterprise plan.
• You want a shared assistant across many channels, not one tightly defined job.
• Your compliance setup allows it, and you’re comfortable with its default search of public channels.

No app is the right call when your team barely uses Slack, because a bot nobody asks is wasted. The same goes for work that is irregular and judgment-heavy, like creative work, complex client relationships or delicate decisions. Automation there tends to cost more upkeep than it saves.

If you’re unsure, run Claude Tag for a while if your plan allows it. Note each time it hits the same wall: missing context, a manual workaround, a question it can’t answer. That list becomes the spec for a custom build, if you need one.

Related Posts

• Internal Knowledge Base AI Integration: What Actually Works for Employee Queries: how to design the retrieval behind a policy or wiki bot.
• Weekly Digest AI Automation for Internal Communications: the scheduled-summary workflow in detail.
• Connecting Claude to Your CRM: What It Takes: the data-source side of a proposal or sales bot.
• Human Review Checkpoints in AI Automation: where a person should approve what the bot produces.

Frequently Asked Questions

What’s the difference between the Claude app in Slack and a custom Claude API integration?

Claude Tag, Anthropic’s current Slack product, is a shared assistant with channel memory, routines and admin-set connectors. Anthropic configures and runs it, and it bills to your Claude organization (source). A custom Claude API integration is your own code. You choose the model, the prompts, the data it reaches and the exact behavior, and you host and maintain it.

How much does a Claude API Slack bot cost to run per month for a small business?

The token cost is often small. In our own worked example above, an HR policy bot answering 200 questions a month comes to about $3.20 on Sonnet 5.5, based on Anthropic’s published prices (source). Your figure depends on how many questions you get and how much context each one carries. Hosting and development usually cost more than tokens, so get them quoted against a written spec.

Do I need a Slack Team or Enterprise plan to use Claude with Slack?

For Claude Tag, the plan that matters is your Claude plan: it needs Claude Team or Enterprise (source). The earlier Claude in Slack app needed a paid Slack plan (source). For a custom bot, Slack says some AI features need a paid Slack plan, and its Developer Program offers a free sandbox for building (source). On Slack’s free plan, only the most recent 90 days of messages are visible (source), which limits what any bot can read.

Can a Claude API bot read our existing Slack history and internal documents?

Yes, with the right scopes. Reading history with conversations.history uses the channels:history, groups:history, im:history and mpim:history scopes, and internal apps get Tier 3 rate limits (source). Documents in Google Docs, Notion or a wiki connect through their own APIs. Claude 4.6 and later models include a 1 million token context window at standard pricing (source). Even so, fetching only what’s relevant keeps answers sharper and costs lower. That retrieval logic is a core part of the build.

Who owns the code and data when we hire someone to build this?

You should. Any contract for custom AI work should say you receive the full source code, own the intellectual property and control where your data is stored. The Claude API is still a third-party service: your prompts pass through Anthropic for inference, under its retention terms (source). Your integration code, your data sources and your prompts are yours if the contract says so. Our own terms are on the AI integration page: no lock-in, and full ownership of whatever is built.

What happens if Claude gives a wrong answer and there’s no human reviewing it?

Sooner or later it will happen. Language models can misread context or state something wrong with confidence. The fixes are design choices. Tell the bot what it must not answer as fact. Make it cite the source it used. Use strict tool definitions so writes to other systems match your schema (source). Send anything with real consequences to a person. A bot that says “I’m not sure, here’s what I found, check with the HR lead” is worth more than one that always sounds certain.