Agenvoy is a powerful Go-based AI agent framework designed to streamline the orchestration of multiple Large Language Model (LLM) providers. It supports seamless routing of tasks across providers such as OpenAI, Nvidia, and Codex, utilizing a planner-based dispatch system. Notably, the framework eliminates the need for HTTP communication through direct execution of subagents, enhancing efficiency and responsiveness.

Key Features

• Multi-Agent Orchestration: Agenvoy integrates seven prominent LLM providers into a single interface, enabling efficient task management and execution. Users can invoke subagents directly without HTTP delays, inheriting relevant context from their parent sessions.

• Pluggable Tools and OS-Native Sandboxing: Users can easily extend functionality by adding JSON configurations or scripts, which will be recognized as tools. The platform supports an MCP client adapter that utilizes standard input/output and HTTP/SSE for real-time management. All tools run within a secure OS-native sandbox using Linux bwrap and macOS sandbox-exec, ensuring safe execution of commands and scripts.

• Self-Improving Error Memory: The framework implements ToriiDB to maintain an intelligent error memory that refreshes based on interactions. This feature allows for improved learning over sessions by employing both keyword and semantic searching techniques to avoid repeating previous mistakes.

Architecture Overview

Agenvoy's architecture is streamlined for performance and usability, connecting critical components like the entry point, execution engine, multiple LLM providers, and tools subsystem efficiently:

graph TB
    Entry[Entry · cmd/app/main.go]
    Engine[exec.Run / Execute<br/>≤128 iterations]
    Providers[LLM Providers · 7]
    Pending[Pending Registry<br/>confirm / ask]

    subgraph Tools[Tool Subsystem]
        MCP[MCP Adapter<br/>stdio / HTTP]
        Sub[Subagent · in-process]
        Ext[External CLI · 4 vendors]
        Builtin[Built-in Tools<br/>file / web / api / script]
    end

    Sandbox[Sandbox<br/>bwrap / sandbox-exec]
    Session[Session<br/>ToriiDB / bot.md / status.json]

    Entry --> Engine
    Engine --> Providers
    Engine --> Tools
    Engine <--> Pending
    Engine <--> Session
    Builtin --> Sandbox

For further details and documentation, visit the Agenvoy Wiki.