Brod: An MCP-Based Cloud for Self-Hosted AI Agent Tools

Overview

A new open-source project called Brod is aiming to solve a persistent problem in enterprise AI adoption: where do the tools built by agents actually live? As AI models become more capable of constructing internal utilities like on-call rotas, stock checkers, and form replacements, they require a secure, centralized hosting environment that does not burden each application with its own login system or database management.

Brod positions itself as a self-hosted cloud specifically designed for AI agent-deployed tools. It leverages the Model Context Protocol (MCP) to allow agents to deploy applications directly, providing a unified sign-in and access control layer for all generated tools within an organization.

The Problem with Agent-Built Tools

Modern AI agents can now build functional software components, but deploying them traditionally involves significant operational overhead. Each small tool typically requires its own server instance, user table, and authentication mechanism. This fragmentation creates security risks and administrative burdens for IT teams.

Brod consolidates these deployments into a single infrastructure layer. When an agent deploys an application via MCP, it receives a secure address that is gated behind the company’s existing identity provider. This ensures that sensitive internal tools are accessible only to authorized personnel without requiring individual account creation for every new utility.

Core Architecture and Deployment

The platform is built on a Rust-based server architecture and utilizes Docker Compose for simplified deployment. A typical installation involves cloning the repository, configuring environment variables, and starting the service, which initializes a control plane, a PostgreSQL database, and a runtime environment based on Cloudflare’s workerd.

The system supports both static and server-side applications. Server apps are written as JavaScript fetch handlers and come with built-in SQLite databases, key-value stores, and file storage. This allows agents to create complex, data-driven tools without managing external infrastructure dependencies.

flowchart LR
  agent["Agent<br/>(Claude Code, Cursor, ...)"] -- "MCP over HTTP<br/>/mcp" --> brod
  people["People's browsers"] -- "app.workspace.APPS_DOMAIN" --> brod
  subgraph server["Your server (docker compose)"]
    brod["brod<br/>control plane, sign-in,<br/>gateway, admin UI"]
    db[("Postgres<br/>workspaces, people,<br/>apps, versions")]
    files[("brod-data<br/>app files, secrets")]
    runtime["runtime<br/>workerd: apps' server code,<br/>one SQLite per app"]
    brod --- db
    brod --- files
    brod -- "server-app requests" --> runtime
  end

Key Features and Security

Brod introduces a workspace-based model where each team or company operates in an isolated environment. Access control is managed through a unified sign-in system, supporting email magic links or Google OAuth, eliminating the need for custom user tables in deployed applications.

Roles and Access Control

The platform offers granular permission settings through workspace roles such as admin, builder, member, and guest. Applications themselves can be configured as private, shared with specific individuals or groups, open to the entire workspace, password-protected, or public. This flexibility allows organizations to balance openness with security depending on the sensitivity of the tool.

"Apps read who is signed in, with their role and groups, and keep no user table of their own."

Deployment Lifecycle

Agents interact with Brod through 13 distinct MCP tools, enabling them to deploy applications, read logs, manage secrets, and configure access. The platform supports a staging workflow where updates are first pushed to a staging address. Publishing and rolling back changes involve moving a pointer, ensuring that updates take effect on the next request without downtime.

Security is reinforced through encrypted secrets that are readable only by the application’s server code. Additionally, concurrent edits by multiple agents are prevented to avoid silent overwrites of work.

Integration with AI Agents

Brod is designed to integrate seamlessly with popular AI coding assistants. It supports configuration for Claude Code, Claude Desktop, and Cursor through standard MCP connection methods. Users can add Brod as an MCP server by providing a generated token and the local or remote endpoint URL.

Once connected, agents can deploy example applications directly from the command line. For instance, an agent can be instructed to deploy a hello-team example as an on-call rota, after which it provides the resulting application address. This workflow enables non-technical users to access agent-built tools through a familiar browser interface while maintaining strict access controls.

Self-Hosted and Hosted Options

While this project focuses on self-hosted deployments, a hosted version is available at brod.sh. Both versions share the same underlying codebase, differing primarily in operational management and feature availability. The hosted service includes marketplace integrations and automated abuse checks, whereas self-hosted installations offer full control over data and infrastructure.

The self-hosted edition is licensed under AGPL-3.0 and imposes no usage limits. Documentation covers configuration, production readiness, security considerations, and architecture details to support enterprise adoption.