jin-bo/agentao

What it solves

Agentao provides a governed runtime for AI agents, addressing the unpredictability and lack of control when deploying agents in local environments. It focuses on privacy, local-first operation, and strict governance to ensure agents act with consent and transparency.

How it works

It functions as an embeddable Python harness that manages the lifecycle of an agent. It implements three core pillars:

• Constraint: Uses permission modes (read-only, workspace-write, full-access, and plan mode) and macOS sandboxing to prevent unauthorized actions.
• Connectivity: Integrates with the Model Context Protocol (MCP) and Agent Context Protocol (ACP) to connect agents to external tools and other agents via JSON-RPC.
• Observability: Provides real-time reasoning chains, streaming tool outputs, and JSONL session recordings for full auditability.

Who it’s for

• Embedding Developers: Those building AI agents into existing Python applications.
• CLI Users: Developers who want a governed, terminal-based AI assistant with persistent memory and skill management.
• Coding Agents: Other AI agents (like Claude Code) that need a standardized way to embed Agentao into a project.

Highlights

• Governed Runtime: Four distinct permission modes and tool confirmation to ensure safety.
• Persistent Memory: SQLite-backed memory with user and project scopes and automatic recall.
• Extensible Skills: Support for installing skills from GitHub and a /crystallize workflow for creating new ones.
• Multi-Agent Support: Built-in personas like codebase-investigator and the ability to define custom agents via markdown files.
• MCP/ACP Support: Native integration with the Model Context Protocol for tool discovery and connectivity.