Brickr Agent

> Give an external AI client controlled access to build in Brickr without sharing your account session.

Each Brickr account can create one personal Agent identity. The Agent has its own name, color, workspace grants, and MCP connections. It is not a workspace member until access to a specific workspace has been approved.

Open the user menu in the dashboard sidebar, select Brickr Agent, or go directly to /dashboard/agent to create and manage it.

Connect with Codex

codex mcp add brickr --url https://mcp.brickr.dev/mcp

Restart Codex or open a new session if Brickr does not appear immediately. The client opens a Brickr connection page when it needs authorization.

Connect with ChatGPT

ChatGPT does not provide a chatgpt mcp add terminal command. Configure Brickr as a custom app in ChatGPT web:

1. Enable Developer mode. Business admins/owners can use Workspace settings > Apps > Create. For Enterprise/Edu, an admin first grants developer access; authorized users then enable it under Settings > Apps > Advanced Settings. 2. Open Settings > Apps > Create. 3. Enter https://mcp.brickr.dev/mcp as the remote MCP endpoint and choose OAuth when authentication is requested. 4. Select Scan Tools and complete the Brickr authorization page. 5. Create the app, then enable or select Brickr in a chat.

Full MCP write actions currently require ChatGPT Business, Enterprise, or Edu on the web. ChatGPT Pro developer mode supports read/fetch connectors, which is not sufficient for asking the Brickr Agent to create or edit APIs. Availability can change; see OpenAI's Developer mode and MCP apps documentation.

Connect with Claude or another client

Use a client surface that supports remote Streamable HTTP MCP servers and OAuth. Add a remote custom connector named Brickr with this endpoint:

https://mcp.brickr.dev/mcp

Client menus and commands vary. Follow that client's current MCP documentation instead of translating the Codex command into a guessed command.

Complete the Brickr connection

The client opens a Brickr connection page. Sign in, review the personal Agent shown on the page, and select Connect Agent. Brickr exchanges the one-time connection code for an opaque Agent session. The MCP server never receives your Brickr account JWT or account email.

Select and approve a workspace

The MCP client does not receive access to every workspace automatically:

1. Ask the client to list your available workspaces. 2. Choose the workspace in which it should work. 3. If no matching grant exists, the Agent returns an approval link. 4. The Brickr dashboard shows the Agent identity, workspace, and exact requested scopes. 5. Approve or deny the request. 6. The MCP client waits for the approval and selects the workspace automatically when possible. If the callback cannot complete, return to the client and confirm that approval is finished.

The approving member must be allowed to manage AI Agents. Brickr also limits the requested Agent scopes to the permissions that member already has. An Agent can never use a grant to exceed the approving member's workspace rights.

Workspace scopes

| Scope | Allows | |-------|--------| | workspace.read | Read basic workspace information | | assets.read | List and inspect workspace assets | | data.read | Read database and form data | | data.write | Create or change workspace data | | builder.read | Read route and function graphs | | builder.write | Create and edit Builder resources | | routes.test | Execute DEV route tests | | routes.deploy | Deploy routes | | secrets.metadata | Read secret names and metadata, never values | | secrets.write | Create or update workspace secrets | | destructive | Delete supported resources |

Grant only the scopes required for the current job. Read-only research generally needs workspace.read, assets.read, data.read, and builder.read. Building and DEV testing additionally needs builder.write, data.write, and routes.test where applicable.

Isolation and revocation

Each active workspace grant resolves to an isolated Agent principal for that workspace. Every MCP tool call revalidates the session, selected workspace, grant, and required scope. Revoked grants stop workspace access; revoked MCP connections invalidate the associated session and API key.

In Brickr Agent settings you can:

  • Rename or recolor the personal Agent
  • Add, edit, or remove workspace grants
  • Review the scopes for every grant
  • See active MCP clients and their selected workspace
  • Revoke a client connection immediately

Brickr Assistant versus Brickr Agent

| Capability | Brickr Assistant | Personal Brickr Agent | |------------|------------------|-----------------------| | Answer workspace and asset questions | Yes | Yes, with read scopes | | Show safe form aggregates | Yes | Yes, with data.read | | Test the current open DEV route | Yes | Yes, with routes.test | | Build or edit flows | No | Yes, with builder.write | | Deploy routes | No | Yes, with routes.deploy | | Delete resources | No | Only with destructive |

Use Brickr Assistant for information in the Brickr UI. Use the personal Agent when an external MCP client should build or manage workspace resources.

What's next?

| Topic | Description | |-------|-------------| | Brickr Assistant | Understand the read-only in-product chat | | Workspace Members | Review human workspace roles and permissions | | Secrets | Store credentials without exposing their values |