✨ Soothe is an agent-harnessing framework—an Agentic OS that pushes humans out of the execution loop. Built on LangChain / DeepAgents, it adds a persistent agentic loop and goal engine that maintains context across sessions, sustains long-running goals, coordinates multiple objectives, and autonomously steers complex tasks. Shift from human-in-the-loop to agent-in-the-loop: define intent, let the system handle execution.
- ✨ Thinks Ahead — Multi-step planning with dynamic adaptation
- 🚀 Acts Autonomously — Research, coding, file ops, plugin automation
- 🧠 Learns & Remembers — Persistent memory across sessions
- 🔒 Stays Secure — Least-privilege, local-first architecture
- 🔌 Extends Easily — Decorator-based plugins, custom tools, subagents
- 🌐 Works Anywhere — Multi-transport daemon (WebSocket, HTTP REST)
| Principle | Description |
|---|---|
| Plan → Execute | Autonomous loop: plan, act, evaluate, adapt |
| Persistent Memory | Resume threads, recall context, track goals |
| Security First | Local execution, least-privilege policies |
| Plugin Architecture | Decorator-based tools, subagents, MCP servers |
| Capability | Features |
|---|---|
| Deep Research | Multi-source web search, academic papers (arXiv, DeepXiv), document analysis |
| Autonomous Execution | Multi-step workflows, file ops, code execution, shell commands |
| Long-Running Ops | Background daemon, thread management, persistent state |
| Custom Plugins | @tool, @subagent, @plugin decorators, MCP server integration |
| Status | Milestone |
|---|---|
| ✅ | Single-Session Autonomy — End-to-end goal execution |
| ✅ | Cross-Thread Continuity — Persistent context across threads |
| ✅ | Multi-Goal Orchestration — Interdependent long-horizon workflows |
| ⏳ | Benchmark Reproduction — Compiler experiment |
Requirements: Python 3.11+
For everyday use you install three PyPI packages: core (soothe), CLI (soothe-cli), and daemon (soothe-daemon). soothe-sdk is pulled in automatically; you only need it directly when building plugins or custom clients.
| Package | Version | Provides |
|---|---|---|
soothe |
 |
Agent runtime, protocols, backends, tools |
soothe-cli |
 |
soothe command |
soothe-daemon |
 |
soothed command (WebSocket, HTTP REST) |
soothe-sdk |
 |
Shared protocol types and decorators |
soothe-community |
 |
Community plugins (separate repo) |
Recommended — full daily stack (core + all optional capability groups + CLI + daemon):
pip install -U 'soothe[all]' soothe-cli soothe-daemonsoothe[all]— research, documents, tabular data, media, GitHub, semantic search, Langfuse, DashScope, and related tool dependenciessoothe-cli— interactive TUI and one-shot prompts (soothe,soothe -p "…")soothe-daemon— long-running server and lifecycle commands (soothed start,soothed doctor)
Minimal — core + CLI only (lighter install; add soothe-daemon when you need background mode):
pip install soothe soothe-cliInstall the daemon later with pip install soothe-daemon. Pick optional capability groups instead of [all] when you want a smaller footprint, for example pip install 'soothe[research,document]' soothe-cli soothe-daemon.
Verify:
soothe --help
soothed doctor1. Configure:
mkdir -p ~/.soothe/config
cp config/config.template.yml ~/.soothe/config/config.yml
export OPENAI_API_KEY="sk-..." # or ANTHROPIC_API_KEY, DASHSCOPE_API_KEY2. Run:
# Start daemon
soothed start
# Interactive TUI
soothe
# Single prompt
soothe -p "Research top 5 Python web frameworks"Commands:
| Command | Description |
|---|---|
soothe |
Interactive TUI |
soothe -p "query" |
Single prompt |
soothed start/stop/status |
Daemon management |
soothed doctor |
Health diagnostics |
| Resource | Description |
|---|---|
| User Guide | End-user usage guide |
| RFCs | Architecture specs |
| CLAUDE.md | AI agent dev guide |
MIT