Page Agent

The Page Agent is an AI assistant built into Bytebase. It lives in a floating chat window that’s available on every page — ask it anything in plain language, and it helps you get it done.

Why we built it

Bytebase has a lot of features, and it’s not always obvious how to accomplish what you’re trying to do — even when the feature you need is already there. You might know what you want (archive old rows, grant a teammate access to a project, fix a failing plan check) but not which page to open, which buttons to click, or which settings control the behavior you’re after. Or maybe you know exactly how to do it — you’d just rather hand it off and review the result instead of clicking through the flow yourself.

The Page Agent is designed for both. It understands how Bytebase works end to end, so it can:

• Show you the way — walk you through a workflow step by step when you want to do it yourself.
• Do it for you — take the actions directly when you’d rather describe the outcome and check the result.

You stay in control: the agent shows each step it takes so you can follow along, and you can stop or redirect it at any time.

Setup

The Page Agent uses the same AI provider configuration as the rest of Bytebase. If you haven’t enabled it yet, follow the steps in AI Assistant to configure your OpenAI, Claude, Gemini, or Azure OpenAI credentials.

Once AI is configured, the Page Agent is available from any page in the Bytebase console. Click the agent toggle in the dashboard header to open the chat window.

How it works

The Page Agent has three layers:

1. Chat window — A floating panel available from any page in the console. It stays open as you move between pages.
2. Agent loop — Runs in your browser. When you ask a question, the agent thinks about what to do, takes an action (reading the page, calling an API, navigating somewhere), observes the result, and repeats until it has a complete answer. Each step is shown in the chat.
3. AI backend — Bytebase routes your conversation to the AI provider configured in your workspace settings. Your credentials never leave your server.

Capabilities

Capability	Description
Read the current page	Understands which page you’re on, what project, database, or issue you’re viewing, and what state the UI is in
Call Bytebase APIs	Queries and updates data directly — listing databases, creating issues, checking deployment status, and more
Navigate	Goes to any page in the console on your behalf
Interact with the UI	Clicks buttons, fills form fields, selects options — useful on pages with unsaved state like editors and creation forms
Discover APIs	Browses the full Bytebase API catalog to find the right endpoint for a task
Load workflow guides	Pulls in step-by-step instructions for common multi-step tasks like running queries, creating schema changes, or granting permissions

Controls and audit

The Page Agent operates within Bytebase’s existing safety mechanisms:

• Uses your permissions. The agent can only perform actions your account is authorized for. If you can’t drop a table, neither can the agent.
• Goes through the normal approval flow. Schema changes the agent creates are regular Bytebase issues — they respect your project’s approval policies, plan checks, and rollout gates. The agent cannot bypass these.
• Audit-logged under your account. Every API call, issue creation, and permission change is recorded in the audit log with your user name.
• Controlled by workspace AI settings. The Page Agent is available only when AI is enabled in your workspace. Disabling AI Assistant in Settings > General turns off the Page Agent along with other AI features.

Use cases

Fixing a plan check error

Your issue failed a SQL review check and you want help understanding and fixing the violation.

You: Help me fix this plan check error.

Agent: (Reads the current page to inspect the issue and SQL review results)

Agent: The SQL review flagged a violation: “Do not apply functions or perform calculations on indexed fields in the WHERE clause.” The problem is in your INSERT…SELECT statement:

INSERT INTO orders_archive (id, customer_id, created_at)
SELECT id, customer_id, created_at
FROM orders
WHERE YEAR(created_at) = 2023;

Wrapping created_at in YEAR() prevents the database from using the index on that column, forcing a full table scan. Here’s a rewrite that keeps the index usable:

INSERT INTO orders_archive (id, customer_id, created_at)
SELECT id, customer_id, created_at
FROM orders
WHERE created_at >= '2023-01-01' AND created_at < '2024-01-01';

Want me to update the SQL in the editor and re-run the check?

Other things you can ask:

• “Why did the SQL review fail on this issue?”
• “The plan check says I have a naming convention error — can you fix it?”
• “Rewrite my migration to pass all the review checks”

Granting permissions

A teammate needs access to a project, and you want to set it up without navigating to settings.

You: Give [e-mail] developer access to the Payment project.

Agent: I’ll add Alice as a project developer. Let me look up the project members API. (Discovers the project IAM API → calls it to add the role binding)

Agent: Done! [e-mail] now has the Project Developer role on the Payment project. She can view databases, create issues, and run queries in SQL Editor.

Other things you can ask:

• “Who has access to the Production environment?”
• “Remove [e-mail] from the HR project”
• “What roles does [e-mail] have?”

Querying data across your workspace

Get a quick overview of resources without clicking through multiple pages.

You: What databases are in the “Payment” project?

Agent: Let me check. (Calls the Bytebase API to list databases in the Payment project)

Agent: The Payment project has 4 databases:

Database	Environment	Engine
payment_prod	Production	PostgreSQL 15
payment_staging	Staging	PostgreSQL 15
payment_dev	Dev	PostgreSQL 15
payment_analytics	Production	ClickHouse

Other things you can ask:

• “How many pending issues are there in the HR project?”
• “Show me all PostgreSQL instances”
• “What SQL review policies are configured for production?”

Tips

• Be specific. “Add a NOT NULL varchar(255) column called display_name to the users table in staging” works better than “change the users table.”
• Ask follow-ups. The agent remembers the full conversation, so “now do the same for production” works after a staging change.
• Let it navigate. You don’t need to be on the right page first — the agent will go where it needs to.
• Watch the tool calls. The chat shows each API call or UI action the agent takes, so you always know what happened.

Important notes

The agent uses AI to plan and execute actions. Always review what it did — issue created, SQL written, permissions granted — before moving on.