> ## Documentation Index
> Fetch the complete documentation index at: https://chatbase.co/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Backstage: manage your AI agent with an AI assistant

> Backstage is an AI-powered operations center that lets users manage their AI agent through natural language conversation, instead of navigating multiple settings pages manually. Think of it as a copilot for AI agent management: users describe what they want in plain English, and Backstage handles the rest.

Backstage is an AI assistant built into the Chatbase dashboard that lets you manage your AI agent by chatting with it. Instead of opening settings pages one by one, you describe what you want, such as "why are customers unhappy this week?" or "make the tone friendlier", and Backstage reads your data, proposes changes, and applies them once you approve. It's available on every plan, on the web, and from Slack and iMessage.

## What Can Users Do With It?

### Analytics & Insights

* Ask questions like "What are my top customer topics this month?" or "Show me conversation trends for the last 3 months"
* View interactive charts (bar, line, area, pie) generated in real time
* Check credit usage breakdowns
* Identify knowledge gaps: questions the AI agent couldn't answer

### Training & Knowledge Base

* Add, update, or delete knowledge base sources (text, Q\&A pairs, URLs)
* Upload files (PDFs, Word docs, spreadsheets, archives) to extract training data from
* Add, edit, or delete sources; each change trains on its own
* Toggle auto-resync on or off

### AI Agent Configuration

* Update the AI agent's system instructions or channel-specific instructions
* Change the AI model or temperature
* Modify the Chat Bubble appearance (colors, header, button styles, themes)

### Helpdesk (Read-Only)

* Query tickets by status, assignee, channel, or date
* See each ticket's tags, custom fields, and priority
* Search tickets using text or semantic search
* Export up to 10,000 tickets to analyze them in bulk. For example, upload a file that maps your tags to topics and ask for tickets per topic per week
* View AI agent performance metrics, resolution analytics, and workload distribution
* Review Helpdesk settings and saved views

### Actions & Integrations

* Create, update, enable/disable, or delete Actions (Slack notifications, webhooks, ticket creation, etc.)
* Build [Custom Actions](#building-custom-actions) from start to finish, including testing them against your real API
* Check integration status and trigger OAuth connections for new integrations

### Deployment

* Check which channels are active (Chat Bubble, WhatsApp, Slack, etc.)
* Turn the Chat Bubble, Help Page, Center Stage, and email channels on or off. The channel must already be set up. Connect and manage WhatsApp, Messenger, and Instagram from the AI agent's [**Channels**](https://chatbase.co/docs/user-guides/chatbot/channels.md) tab.

### Guardrails and conversations

* Change [spam detection](https://chatbase.co/docs/user-guides/chatbot/guardrails.md#spam-detection) and [country blocking](https://chatbase.co/docs/user-guides/chatbot/guardrails.md#country-blocking). Both need your approval.
* Pause or resume a conversation

### Versions

* Publish, discard, or restore [versions](https://chatbase.co/docs/user-guides/chatbot/versions.md) of your AI agent

### Built-in Skills (Multi-Step Workflows)

Users don't need to invoke these explicitly. The AI uses them automatically when relevant:

* Weekly insights report
* Tone/instruction updates with safe diff preview
* Drafting Q\&A from knowledge gaps
* New AI agent onboarding walkthrough
* File/archive processing into training data
* Helpdesk triage, issue review, performance review, and settings audit

***

## How Does Approval Work?

By default, any mutative action (updating instructions, adding sources, deleting an action, etc.) triggers a confirmation UI where the user can review each change and approve or reject it individually before anything is applied. Text changes show a side-by-side diff; settings changes show a before/after table.

To let Backstage make some changes without asking each time, turn on [auto mode](#auto-mode).

***

## Auto Mode

With auto mode on, Backstage can change your AI agent without asking you to approve each edit. Changes you ask for are made right away, and improvements Backstage proposes are tested before they're published.

### Turn on auto mode

**1. Open Backstage**

Open Backstage on any AI agent.

**2. Choose what Backstage may change**

Click the gear icon next to **Auto mode**, select the areas Backstage may change without asking, and click **Save**.

**3. Turn it on**

Turn on the **Auto mode** toggle.

Auto mode applies to the current chat only. You can turn it off at any time with the toggle, or by telling Backstage in the chat.

### What auto mode changes

You choose which of these areas Backstage may change without asking:

* Instructions and model
* Appearance
* [Procedures](https://chatbase.co/docs/user-guides/chatbot/procedures/procedures-overview.md)
* [Widgets](https://chatbase.co/docs/developer-guides/widgets/overview.md)
* Added knowledge (text and Q\&A)

When you explicitly ask for a change in one of the selected areas, Backstage makes it right away, without running tests. There's no approval card for each edit, and no re-approving when you tweak the same change again.

Anything outside the selected areas, such as editing actions or deleting sources, still asks for your approval first.

> [!NOTE]
>
> Changes you make still wait for you to publish. Backstage asks you once before they go live.

### How Backstage fixes a problem

When Backstage sees your AI agent answering wrong, whether in conversations, a knowledge gap, or a failed [test](https://chatbase.co/docs/user-guides/chatbot/tests.md), it proposes an improvement. Unlike changes you ask for, improvements are always tested before they go live. Once you approve it, Backstage:

**1. Measures the problem**

Runs test conversations against the live AI agent to measure the problem.

**2. Makes the change in its own draft**

Applies the fix in a separate draft. Your draft and the live AI agent stay untouched.

**3. Tests the change**

Runs the same tests on the changed AI agent.

**4. Publishes or rolls back**

If the change passes without making your saved tests worse, Backstage publishes a new version. If it fails, Backstage rolls it back and tries again, up to **3 times**, then tells you what it tried.

You can watch the tests run in the chat and cancel at any time.

> [!NOTE]
>
> While an improvement is running, **Publish** and **Restore** are paused on the dashboard, so two changes never collide.

### Auto mode on Slack

Auto mode works the same way from [Slack](#connect-slack). Approvals and results arrive in the thread, and replying **"stop"** cancels a running improvement.

***

## Scheduled Audits

Backstage can audit your AI agent on a schedule and send you a report, so you hear about problems without having to open Backstage and ask. Reports can run **daily**, **on weekdays**, or **weekly**.

### Set up an audit

Ask Backstage in your own words. For example:

> Send me a daily audit at 9am. Lead with refund requests, and flag anyone asking about shipping to Canada.

Backstage asks for anything it still needs, such as where to send the report, which day to use for a weekly report, or your city (on Slack or iMessage) so it can send the report at the right local time. It then shows you the schedule for approval and sets it up.

Anything you ask it to focus on becomes a standing instruction for every report. You can change the audit later in conversation, for example:

* "Also watch for complaints about late deliveries."
* "Make it weekly."
* "Pause it."

### Where reports arrive

* **Slack or iMessage**, if you've connected one (see [Connect Slack](#connect-slack) and [Connect iMessage](#connect-imessage))
* Otherwise, a **Backstage session** in the dashboard, along with an email letting you know a new report is ready

Replying to a report continues the conversation, so you can say things like **"fix that"** or **"tell me more about #2"**.

### What's in a report

> [accordion] Needs you
>
> Problems only you can fix, with what to do about each one.

> [accordion] Problems
>
> What's going wrong, with the numbers or conversations behind it.

> [accordion] Working well
>
> Real improvements since the last report.

> [accordion] Since last report
>
> What got fixed, what's still open, and what's new.

> [accordion] Suggestions
>
> At most two suggestions, smallest fix first: a Q\&A, a source fix, an instructions change, then an Action. If you say yes, Backstage builds the fix for you (and asks for your approval before applying it).

Every audit goes through the same checklist:

* **Conversation quality**: thumbs-down ratings, escalations, failed Actions, and customers asking for a human
* **Knowledge**: gaps and conflicting sources
* **Instructions**: your AI agent's general and channel-specific instructions
* **Deployment and training status**
* **Procedures and Actions**
* **Helpdesk queue**

Backstage reviews every conversation in the reporting window, not a sample. If nothing needs attention, the report is a one-line all-clear.

> [!TIP]
>
> For a one-time audit, just ask **"audit my agent"**. Backstage runs the same checklist over the last 7 days.

### Good to know

* Each user can have up to **3 scheduled reports per AI agent**.
* Scheduled runs are **read-only**. Nothing changes until you reply and approve a fix.
* Reports don't use message credits and don't count against your daily Backstage message limit.
* HIPAA accounts don't receive email notifications. Reports still appear in Backstage.

***

## Building Custom Actions

Backstage can build a [Custom Action](https://chatbase.co/docs/user-guides/chatbot/actions/custom-action.md) entirely in conversation: the endpoint, headers, query parameters, body, and the details your AI agent asks customers for. It then tests the endpoint against your real API and turns the action on.

### Build an Action

Describe what your AI agent should be able to do, and include links to API docs if there are any. For example:

> I want my agent to tell people the current weather in their city. Here are the docs: [https://open-meteo.com/en/docs](https://open-meteo.com/en/docs)

Backstage reads the docs, drafts the Action, and asks for one approval that covers the whole build: the configuration, the test call, and turning the Action on. If your request leaves a real choice open, such as which endpoint to use or whether the action should change data, Backstage asks you one question first.

> [!NOTE]
>
> Backstage works out how the API expects its inputs. In the weather example, it notices that the API takes coordinates rather than city names, so it builds the Action with latitude and longitude as inputs.

### API keys stay out of the chat

When an endpoint needs an API key, Backstage shows a masked input. The value goes straight from your browser to the Action, and Backstage never sees it. When Backstage reads the Action back later, any saved key appears as `[set]`.

The form also shows which Action the key is for and the host it will be sent to, so you can check both before pasting a live key.

### Testing against your real endpoint

Nothing is sent to your API until you click **Run**. Before that, the test card shows the exact request that will be sent, and warns you when the method can change data (`POST`, `PUT`, `PATCH`, or `DELETE`).

Backstage only receives the shape of the response (field names, types, and short samples), never the full body. Those fields become the data your AI agent can see when it calls the Action. You can narrow them later from the Action's [Data Access](https://chatbase.co/docs/user-guides/chatbot/actions/custom-action.md#data-access) settings in the dashboard.

### Spotting missing Actions

When Backstage reviews your conversations, it can notice that customers keep asking for something only an Action could answer, and offer to build that Action. Many gaps are better solved with a Q\&A or a source fix, and in those cases Backstage suggests that instead of an action.

### Good to know

* If the request uses a `{{variable}}` that your AI agent is never told to collect, Backstage points it out.
* On Slack, the **Run** button works right in the thread. Entering an API key opens Backstage on the web.

***

## Plan Access & Daily Message Limits

Backstage is available on every plan. What changes by plan is the daily message limit:

| Plan | Daily messages per AI agent |
| - | - |
| Basic | 10 |
| Business | 100 |
| Enterprise | 100 |
| Agency ([Experts Program](https://chatbase.co/docs/user-guides/chatbase-experts/experts-program.md)) | 100 |

### How the daily limit works

* **It's per AI agent.** Every teammate shares the same daily count for an AI agent, and messages sent from the web, Slack, and iMessage all count toward it.
* **It resets at 00:00 UTC** every day.
* **It's separate from message credits.** Chatting with Backstage doesn't use your workspace's message credits.
* **A second rate limit** of 10 messages per 60 seconds per user prevents bursts.

When you reach the limit on the web, Backstage shows a message like this one (example for a Basic plan):

> You've reached your daily Backstage limit (10/10 messages). Upgrade your plan for more messages.

On Slack and iMessage, the reply is:

> Daily Backstage message limit reached. Upgrade your plan for more messages.

Each request you send counts, so clicking an approval card may also use a message.

### What uses credits

| Backstage activity | What it uses |
| - | - |
| Chatting with Backstage | The daily Backstage limit only |
| Test conversations Backstage runs (for example, when it [fixes a problem](#how-backstage-fixes-a-problem) or tests a change) | Your monthly [test credits](https://chatbase.co/docs/user-guides/chatbot/tests.md#test-credits) first, then message credits once they run out |
| [Scheduled audit](#scheduled-audits) reports | Nothing. They don't use credits or the daily limit. |

[Auto mode](#auto-mode) doesn't change billing. It only decides which changes Backstage can make without asking. If both test credits and message credits are used up, Backstage can't run tests and replies "The account is out of test credits."

***

## Backstage on Slack & iMessage

You can now talk to your AI agent's Backstage directly from Slack or iMessage, no need to open the dashboard. Ask questions, review analytics, manage your AI agent, and approve changes from the apps you already use.

### What you can do from a DM or a text

> [card] Analyze your AI agent
>
> Ask questions like **"How did my AI agent do this week?"** or **"What did users ask about most?"** Charts and reports are delivered directly in the conversation.

> [card] Update your AI agent
>
> Change models, instructions, chat interface, branding, and other settings using natural language. Changes require your approval before they're applied, unless [auto mode](#auto-mode) covers them.

> [card] Manage sources & Actions
>
> Inspect, update, and retrain knowledge sources, and manage Actions without opening the dashboard.

> [card] Approve safely
>
> Review exactly what's changing before approving. Outside of auto mode, nothing is applied without your explicit confirmation.

> [card] Analyze files (Slack)
>
> Upload a CSV or other supported file and ask questions about its contents directly from Slack.

> [card] Continue anywhere
>
> Start on Slack or iMessage and continue the same Backstage session on the web with full history, richer approvals, and interactive charts.

### Approving changes

Every change proposed by Backstage requires your explicit approval, unless [auto mode](#auto-mode) covers it.

> [tab] Slack
>
> Review the proposed changes and approve or reject them directly from Slack.
>
> Each proposal includes:
>
> <ul>
>   <li>A detailed diff of what will change</li>
>   <li>**Approve** and **Reject** controls for every change</li>
> </ul>

> [tab] iMessage
>
> React with:
>
> <ul>
>   <li>👍 to approve</li>
>   <li>👎 to reject</li>
> </ul>

You can also reply with messages like **"yes"**, **"approve"**, **"no"**, or **"reject"**, or equivalent phrases in any language.

> [!NOTE]
>
> If you send another request while an approval is pending, the previous proposal is automatically superseded. You won't accidentally approve an outdated change.

### Connect Slack

1. Open your **Dashboard**.
2. Select your AI agent.
3. Open **Backstage**.
4. Click **Controllers** in the top-right corner.
5. Expand **Slack**.
6. Click **Add to Slack**. You only need to do this once per workspace.
7. Copy the generated link message.
8. Send it as a DM to the **Backstage** Slack app.
9. Once you receive **Connected ✓**, you're ready to chat.

> [!NOTE]
>
> Each Slack thread is its own Backstage session.
>
> If your Slack account has access to multiple AI agents, starting a new thread will ask which AI agent you'd like to use.

### Connect iMessage

1. Open **Controllers** from the Backstage page.
2. Expand **iMessage**.
3. Text the generated link message to the displayed phone number, or scan the QR code to open Messages with the number and message pre-filled.
4. Once you receive **Connected ✓**, you can start chatting.

> [!NOTE]
>
> Your phone maintains one active conversation.
>
> To switch AI agents, send another AI agent's generated link message. This starts a new Backstage session for that AI agent.

## Good to know

> [accordion] Link messages
>
> Link messages are **single-use** and expire after **10 minutes**.

> [accordion] Permissions
>
> Connections are personal. They inherit your Chatbase permissions and can be revoked at any time from the **Controllers** dialog.

> [accordion] Attachments
>
> Slack supports uploading files for analysis.
>
> Backstage can't read attachments sent over iMessage. It replies that it can't open the file, and still answers any text you sent with it. A message that's only an attachment doesn't count toward your daily limit. To share a file, use Backstage on the web or in Slack.

> [accordion] Web experience
>
> Large analyses, interactive charts, and long approval flows include a link back to the corresponding Backstage session in the dashboard.

> [!WARNING]
>
> Nothing is changed without your explicit approval, except changes you've allowed in [auto mode](#auto-mode).

***

## Limitations

* **Message length**: 10,000 characters max per message
* **File uploads**: Max 50 MB per file, up to 5 files per message
* **Analytics range**: Queries can cover a maximum of 6 months
* **Ticket exports**: Up to 10,000 tickets per export
* **Knowledge base source size**: Max 1 MB per text source or Q\&A entry
* **Helpdesk is read-only**: Backstage can surface Helpdesk analytics and tickets, but it **cannot** modify ticket status, assign tickets, or send messages to customers
* **File tools are agent-focused**: Backstage will only process uploaded files for agent-related tasks (improving knowledge base, instructions, analytics). It won't do general-purpose tasks like summarizing a random document or solving homework

### What Backstage can't change

Backstage can't do these. Use the dashboard instead:

* Change the [Rate limit](https://chatbase.co/docs/user-guides/chatbot/guardrails.md#rate-limit) on the **Guardrails** page
* Add, edit, or delete [contacts](https://chatbase.co/docs/user-guides/chatbot/contacts/contacts-overview.md) or contact custom attributes. The custom fields Backstage manages are Helpdesk ticket fields.
* Change the credits limit, notifications, email, Helpdesk, custom domains, or security [settings](https://chatbase.co/docs/user-guides/chatbot/settings.md)
* Delete the AI agent or delete conversations
* Change billing or plans. Backstage can read your credit usage, but can't buy credits or change plans.
* Invite members or change roles
* Connect WhatsApp, Messenger, or Instagram

Backstage acts with your own Chatbase permissions. If your role doesn't allow a change, Backstage replies "Insufficient permissions".

### HIPAA workspaces

Backstage works on HIPAA workspaces, with these limits:

* It can't run full-text searches across conversations.
* Scheduled audit reports aren't emailed. They appear in Backstage only.
* It can't turn on lead email notifications for Actions.

***

## Frequently asked questions

> [accordion] Does Backstage use message credits?
>
> No. Chatting with Backstage only counts toward its own daily message limit, which is per AI agent and resets at 00:00 UTC. The exception is test conversations Backstage runs: they use your monthly [test credits](https://chatbase.co/docs/user-guides/chatbot/tests.md#test-credits) first, then message credits once those run out.

> [accordion] Which plans include Backstage?
>
> Every plan. The daily limit is 10 messages on Basic and 100 on Business, Enterprise, and Agency.

> [accordion] When does the daily Backstage limit reset?
>
> At 00:00 UTC every day. The limit is shared by everyone on your team for that AI agent, across the web, Slack, and iMessage.

> [accordion] Do scheduled audits use credits?
>
> No. Scheduled audit reports don't use message credits or your daily Backstage limit.

> [accordion] Does auto mode cost more?
>
> No. [Auto mode](#auto-mode) only decides which changes Backstage can make without asking. Tests that Backstage runs to check an improvement use test credits first, with or without auto mode.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.