Summary

AgentSphere is an AI Agent orchestration platform enabling closed-loop automation (Perception → Planning → Execution → Feedback). It utilizes an LLM-driven decision engine combined with diverse capabilities like built-in tools, browser automation, CLI execution, and the MCP protocol. This framework targets developers and researchers building sophisticated AI agents, offering a robust system for complex task execution.

How It Works

The platform centers on a ReAct (Reasoning + Acting) execution loop managed by SessionRunner, orchestrating Plan → Act → Observe → Learn cycles. Its Capability Layer integrates various functionalities: MCP for service connections (Jira, GitHub), Java SPI tools, Chrome extension bridge for browser interaction, CLI execution, and composite skills. AgentSphere features a multi-level memory system with persistent PostgreSQL storage, runtime caching, and token budget-based context compaction for efficient long-term conversation management. LLM integration supports multiple providers with automatic routing and fallback.

Quick Start & Requirements

Setup involves building the Java 21 backend (Maven), running a Node.js frontend (React), and loading a Chrome extension. Prerequisites include Java 21, Maven, PostgreSQL, Redis, Node.js, and Chrome. Detailed instructions are in QUICK_START.md.

Highlighted Details

• Supports multiple LLM providers (OpenAI, DeepSeek, Zhipu AI, LiteLLM) with dynamic routing and fallback.
• Advanced multi-level memory system with context assembly, token budget-based compaction, and tool result compression.
• Deep browser integration via a Chrome Extension bridge for DOM manipulation and visual feedback.
• MCP protocol enables seamless connection to external services like Jira, GitHub, and Slack.
• Comprehensive observability via SSE events, Run Activity API, and Session Panel.

Maintenance & Community

The README provides no specific details on maintainers, community channels (e.g., Discord, Slack), or a public roadmap.

Licensing & Compatibility

Released under the permissive MIT License, allowing broad adoption, modification, and commercial use.

Limitations & Caveats

The documentation lacks explicit mention of limitations or known issues. The multi-component setup (Java backend, Node.js frontend, database, extension) presents a moderate entry barrier. The "primary closed loop" phrasing may suggest ongoing development or incomplete features.