psi-agent 0.0.1a20260513

A portable and componentized agent framework

psi-agent

中文 | English

A portable and componentized agent framework.

Introduction

psi-agent is built on two core principles:

• Portability: Copy just the workspace directory to migrate your agent completely.
• Componentization: Agents are assembled from independent components that communicate via Unix sockets, ensuring loose coupling.

Installation

Using uvx (Recommended)

For quick one-off usage without cloning the repository:

uvx psi-agent <component> <subcommand> [options...]
# e.g., uvx psi-agent session --workspace ./workspace ...

Using pip

pip install psi-agent

Using uv

uv add psi-agent

Quick Start

1. Create a workspace directory with your tools and skills (or use an example from examples/):

mkdir -p workspace/{tools,skills,systems}

See examples/a-simple-bash-only-workspace/ for a minimal example, or examples/an-openclaw-like-workspace/ for a more complete setup.

2. Start the session:

psi-agent session \
  --workspace ./workspace \
  --channel-socket ./channel.sock \
  --ai-socket ./ai.sock

3. Start an AI provider (e.g., OpenAI-compatible):

psi-agent ai openai-completions \
  --session-socket ./ai.sock \
  --model <model-name> \
  --api-key <your-api-key> \
  --base-url <provider-api-url>  # e.g., https://openrouter.ai/api/v1

4. Start a channel to interact with your agent session:

psi-agent channel repl --session-socket ./channel.sock

CLI Interfaces

psi-agent provides two CLI interfaces:

1. Subcommand interface (preferred): psi-agent <component> <subcommand>

   • Works with uvx without cloning the repository
   • Single entry point, easier to discover
   • Examples: psi-agent ai openai-completions, psi-agent channel repl

2. Standalone commands: psi-<component>-<subcommand>

   • Shorter to type
   • Useful in scripts
   • Examples: psi-ai-openai-completions, psi-channel-repl

Both interfaces are functionally identical. The subcommand interface is recommended for general use.

Components

psi-agent consists of four component types:

Component	Purpose
psi-ai-*	LLM provider adapters (OpenAI, Anthropic, etc.)
psi-session	Agent runtime loop, handles messages and tool calls
psi-channel-*	Message channels (REPL, Telegram, Feishu, etc.)
psi-workspace-*	Workspace packaging and mounting tools

Available Commands

After installation, these commands are available:

Subcommand format (preferred):

• psi-agent ai openai-completions - OpenAI-compatible completion server
• psi-agent ai anthropic-messages - Anthropic messages server
• psi-agent session - Agent session runtime
• psi-agent channel repl - Interactive REPL interface
• psi-agent channel cli - CLI channel interface
• psi-agent channel telegram - Telegram bot channel
• psi-agent workspace pack - Package workspace to squashfs
• psi-agent workspace unpack - Extract squashfs to directory
• psi-agent workspace mount - Mount squashfs as overlayfs
• psi-agent workspace umount - Unmount workspace
• psi-agent workspace snapshot - Create workspace snapshot

Standalone format (also available):

• psi-ai-openai-completions
• psi-ai-anthropic-messages
• psi-session
• psi-channel-repl
• psi-channel-cli
• psi-channel-telegram
• psi-workspace-pack
• psi-workspace-unpack
• psi-workspace-mount
• psi-workspace-umount
• psi-workspace-snapshot

Documentation

For detailed documentation including workspace structure, tool development, and configuration, see CLAUDE.md.

License

GNU Affero General Public License v3.0 - see LICENSE.md for details.