Open Agent Platform Documentation

Open Agent Platform is a citizen developer platform designed for building, prototyping, and using agents. It provides a web-based interface for managing and interacting with LangGraph agents.

Key capabilities include:

• Agent Management: Building and configuring agents via an intuitive UI.
• RAG Integration: Native support for Retrieval Augmented Generation using LangConnect.
• MCP Tools: Extending agent capabilities by connecting to Model Context Protocol (MCP) servers.
• Agent Supervision: Orchestrating multiple agents through an Agent Supervisor.
• Authentication: Integrated access control and authentication.

An Agent is a custom assistant that is mapped to a specific graph within a LangGraph deployment. It serves as a personalized interface for a particular workflow or logic graph.

Key characteristics of an Agent include:

• Identity: Each agent has a user-provided name and description.
• Configurability: Agents use a custom configuration object to define and personalize their behavior. This object allows you to specify parameters such as the model, system prompt, and available tools used by the underlying graph.

Agents require API keys for models and tools. In Open Agent Platform, users can configure these keys via the settings page.

Key behaviors:

• Runtime Config: Keys set by users are passed to agents under the apiKeys field in the runtime configuration.
• Supported Keys: The demo agents support OPENAI_API_KEY, ANTHROPIC_API_KEY, GOOGLE_API_KEY, and TAVILY_API_KEY. This is extendable for custom agents.
• Fallback Logic: Developers can choose to require user-provided keys or fall back to environment variables set in the LangGraph Platform deployment itself.

Note on Demo Instance:

• Tools Agent and Supervisor Agent: Use user keys first, then fall back to LangChain team keys.
• Deep Researcher: Requires users to set their own API keys in OAP.

Authentication for the RAG server is handled via Supabase. The Open Agent Platform passes the Supabase JWT in the Authorization header of every request to the LangConnect server.

The LangConnect server extracts and verifies this token with Supabase to retrieve a user ID. This ensures that users can only access collections that belong to them. Note that sharing collections between different users is not currently supported.

When building for OAP, you can implement one of three agent architectures:

1. Standard Agents: Single-purpose agents designed to handle specific, isolated tasks.
2. Tools Agents: Agents capable of accessing external tools via the MCP protocol.
3. Supervisor Agents: Orchestration agents that coordinate and manage multiple other agents.

Backend Server

OAP does not require a standalone backend server for the web app to function, provided you have added deployments from the LangGraph Platform to your instance.

RAG Features

To use Retrieval Augmented Generation (RAG) features, you must have a LangConnect server running independently. The LangConnect RAG server is designed to be hosted separately from your LangGraph deployments.

Authentication for the LangConnect RAG server follows the same pattern as LangGraph authentication.

1. Client Request: Pass the Supabase JWT in the Authorization header.
2. Server Verification: The RAG server extracts and verifies the token with Supabase to obtain a user ID.
3. Access Control: The user ID is used to ensure users can only access their own collections.

Note: Sharing collections between users is not currently supported.