Alcyoneus OS is a production-grade Python framework for building intelligent agents and orchestrating multi-agent state-graph workflows. It provides an LLM-agnostic workflow engine with built-in persistence, streaming, human-in-the-loop, guardrails, multi-agent orchestration, sandboxing, and enterprise-grade observability.
import alcyoneus as alc
from alcyoneus.core import StateGraph, Agent, ToolNode, START, END
from alcyoneus.storage.checkpointer import InMemoryCheckpointer
from alcyoneus.prebuilt.tools import safe_calculator, google_web_search
# 1. Define a tool
def get_weather(location: str) -> str:
return f"It's sunny in {location}"
# 2. Build graph with Agent + tools
graph = StateGraph()
graph.add_node("agent", Agent(model="gemini/gemini-2.5-flash", tool_node="tools"))
graph.add_node("tools", ToolNode([get_weather]))
graph.add_edge("agent", "tools")
graph.add_edge("tools", "agent")
graph.set_entry_point("agent")
# 3. Run with persistence
compiled = graph.compile(checkpointer=InMemoryCheckpointer())
result = compiled.invoke(
{"messages": [{"role": "user", "content": "Weather in NYC?"}]},
config={"thread_id": "demo"}
)
✨ Key Features (100+ Production-Grade Capabilities)
🤖 Agent Engine & Orchestration
| Feature |
Description |
| Agent Class |
Build complete agents in 10-30 lines |
| StateGraph |
LangGraph-inspired engine with nodes, edges, conditional routing |
| ReactAgent |
Reasoning + tool-use loops with ReAct pattern |
| RAGAgent |
Retrieval-augmented generation with vector stores |
| SwarmAgent |
Dynamic multi-agent collaboration |
| SupervisorTeamAgent |
Hierarchical agent coordination |
| PlanActReflectAgent |
Plan → Execute → Reflect cycles |
| StructuredOutputAgent |
Guaranteed JSON schema output |
| AudioAgent |
Real-time audio-to-audio (Gemini Live, OpenAI Realtime) |
🔄 Orchestration & Control Flow
| Feature |
Description |
| StateGraph |
Nodes, edges, conditional routing, subgraphs |
| Dynamic Routing |
Command(goto=..., update=...) for inline routing |
| Subgraph Streaming |
Nested graph execution with CompiledGraph returns |
| MessageGraph |
High-level message-only graph wrapper |
| RemoteGraph/GraphServer |
HTTP/gRPC graph execution & serving |
| Dynamic Inline Routing |
Command(goto=..., update=...) |
| MessageGraph |
High-level message-only wrapper |
| All/ALL Selector |
Catch-all edge routing (*, All, ALL) |
🤖 LLM Providers (100+ Models)
| Provider |
Models |
| OpenAI |
GPT-4o, GPT-4o-mini, o1, o3-mini, Responses API |
| Google |
Gemini 2.5 Flash/Pro, Gemini Live, Vertex AI |
| Anthropic |
Claude 3.5 Sonnet, Haiku, Opus |
| LiteLLM |
100+ models (Ollama, Together, Bedrock, etc.) |
| MultiProvider |
Prefix routing: openai/, gemini/, ollama/, anthropic/ |
| Fallbacks |
Exponential backoff, multi-provider retry |
🛠️ 50+ Built-in Tools
| Category |
Tools |
| Search |
google_web_search, bing_search, brave_search, duckduckgo_search, serpapi_search, tavily_search, exa_search, multi_search (7 providers + dedup) |
| Browser |
browser_navigate, browser_click, browser_fill, browser_extract, browser_screenshot, browser_close |
| Shell |
shell_command (safe, policy-guarded) |
| Code |
code_interpreter, CodeInterpreterTool (sandboxed Python) |
| Files |
file_read, file_write, list_directory, edit_file, file_search (semantic + incremental) |
| Search |
google_web_search, vertex_ai_search |
| Image |
generate_image, dalle_generate, imagen_generate, sdxl_generate, midjourney_generate |
| Code |
code_interpreter, CodeInterpreterTool |
| Shell |
shell_command (policy-guarded), ShellTool |
| Files |
file_read, file_write, list_directory, edit_file (unified diff), file_search |
| Calendar |
calendar_create_event, calendar_update_event, calendar_delete_event, calendar_list_events |
| Scheduler |
Scheduler, schedule_job, cancel_scheduled_job, list_scheduled_jobs |
| Subagents |
start_subagent, SubagentManager, create_handoff_tool |
| Browser |
browser_navigate, click, fill, extract, screenshot, close |
| Memory |
memory_tool, create_memory_preload_node, make_agent_memory_tool |
| Image Gen |
generate_image, dalle_generate, imagen_generate, sdxl_generate, midjourney_generate |
| Code |
code_interpreter, CodeInterpreterTool (sandboxed Python) |
| Shell |
shell_command (policy-guarded), ShellTool (Docker/Local) |
| Files |
file_read, file_write, list_directory, edit_file (unified diff), file_search |
| Search |
7 providers: Google, Bing, Brave, DuckDuckGo, SerpAPI, Tavily, Exa + multi_search |
| Image Gen |
generate_image (unified), dalle, imagen, sdxl, midjourney |
| Calendar |
calendar_create_event, calendar_update_event, calendar_delete_event, calendar_list_events |
| Scheduler |
Scheduler, schedule_job, cancel_scheduled_job, list_scheduled_jobs |
| Subagents |
start_subagent, SubagentManager, create_handoff_tool |
| Browser |
browser_navigate, click, fill, extract, screenshot, close |
| Memory |
memory_tool, create_memory_preload_node, make_agent_memory_tool |
💾 Persistence & Memory (3-Layer Architecture)
| Layer |
Implementation |
| Working State |
AgentState (Pydantic, in-memory) |
| Checkpointer |
InMemoryCheckpointer, PgCheckpointer (PostgreSQL + Redis), SqliteCheckpointer |
| Vector Store |
QdrantStore, Mem0Store, MemoryConfig, AgentMemoryConfig |
| Memory Tools |
memory_tool, create_memory_preload_node, make_agent_memory_tool |
| Compaction |
DynamicCompactionPolicy, ResponsesCompactionSession |
| Session Storage |
SQLite, Redis, MongoDB, SQLAlchemy, Dapr, Encrypted |
🌊 Streaming & Real-time
| Feature |
Description |
| Event Streaming |
astream_events() - 15 event types (graph/node/tool start/end, tool calls, handoffs, interrupts) |
| GraphRunStream v3 |
SSE streaming, heartbeat keep-alive, sync wrapper |
| Granularity |
LOW/PARTIAL/FULL response granularity |
| Realtime Audio |
AudioAgent (Gemini Live, OpenAI Realtime, Azure, Local Whisper+TTS) |
| Barge-in |
Audio interruption with transcript persistence |
| WebRTC Streaming |
RemoteDesktopStreamer for VNC/WebRTC frame broadcast |
🛡️ Security & Guardrails
| Feature |
Description |
| Input Guardrails |
PII detection, prompt injection prevention, length limits |
| Output Guardrails |
JSON schema enforcement, blocked words, length limits |
| Tool Guardrails |
ToolInputGuardrail/ToolOutputGuardrail with allow/deny/rate limits |
| Policy Engine |
9-priority RBAC (allow/deny/ask_user), predicates, tenant scoping |
| RBAC |
9 permissions (Viewer/Dev/Admin/Owner), tenant scoping |
| Auth |
JWT (JWKS), mTLS, Token Introspection (RFC 7662), API Key Manager |
| Secrets |
Vault, AWS/GCP/Azure Key Vault, Composite fallback chain |
🌐 Protocols & Communication
| Protocol |
Implementation |
| A2A |
Server (Starlette, agent cards, streaming), Client (retries, pooling, TLS) |
| ACP |
Client/Server (HTTP/in-memory), agent discovery, task delegation |
| MCP |
Stdio/SSE/WebSocket transports, capability negotiation, 5-min tool cache |
| A2A Server |
Starlette app, agent cards, task streaming, uvicorn |
| ACP |
HTTP/in-memory transports, agent discovery, task delegation |
🛠️ Developer Experience & CLI
| Tool |
Purpose |
alc CLI |
alc graph create/run/visualize/validate, alc agent create, alc tool list/test, alc config, alc deploy docker/helm/k8s |
| Graph Visualizer |
compiled.generate_graph("mermaid" | "graphviz" | "html") - renders in GitHub, Notion, browser |
| Graph Visualizer |
Interactive HTML with browser preview |
| Graph Debug |
alc debug state/replay/trace for debugging checkpoints and traces |
🖥️ Sandboxing & Isolation
| Backend |
Capabilities |
| DockerSandbox |
GPU passthrough, resource limits, volume mounts, PTY |
| K8sSandbox |
Pod lifecycle, exec, resource quotas, GPU |
| FirecrackerSandbox |
Micro-VM, resource isolation |
| LocalSandbox |
Unix PTY, subprocess |
| Computer Use |
X11/Wayland/Headless/VNC/Remote Desktop, Accessibility Bridge, Action Verifier |
| ShellTool |
Docker/Local environments, policy-guarded, workspace scoping |
📊 Observability & Observability
| Feature |
Description |
| OpenTelemetry |
Auto-instrumentation decorators (@trace_graph, @trace_node, @trace_llm, @trace_tool) |
| Prometheus |
15+ metrics (requests, latency, tokens, errors, sessions) |
| Structured Logging |
JSON + W3C traceparent propagation |
| 8 Span Types |
Agent, Generation, Function, Guardrail, Handoff, Custom, Task |
| Tracing |
Nested spans, processors (OTLP, Datadog, Console) |
🧪 Testing & Evaluation
| Tool |
Purpose |
QuickTest |
One-liner single/multi-turn assertions |
TestAgent |
Full simulation with MockLLM, MockToolRegistry |
MockMCPClient |
MCP server mocking |
AgentEvaluator |
LLM-as-judge, trajectory matching, safety, hallucination |
QuickTest |
One-liner single/multi-turn assertions |
TestContext |
Isolated test environments |
UserSimulator |
AI-powered user simulation (personas, error injection) |
BatchSimulator |
Concurrent simulation runs |
🏭 Production Hardening (17 Phases Complete)
| Phase |
Area |
Status |
| 1 |
Sandbox Hardening |
✅ Docker, K8s, Firecracker, RemoteFileSync |
| 2 |
Computer Use |
✅ X11/Wayland/Headless/VNC/Remote Desktop/Accessibility |
| 3 |
Shell Tool |
✅ Docker/Local environments, policy-guarded |
| 4 |
Web Search |
✅ 7 providers + multi_search dedup |
| 5 |
Image Gen |
✅ DALL-E, Imagen, SDXL, Midjourney |
| 5b |
File Search |
✅ Semantic index, incremental, multi-repo |
| 6 |
Realtime Audio |
✅ OpenAI/Azure Realtime, Local Whisper+TTS |
| 7 |
Distributed Sync |
✅ yjs/Automerge CRDT, mDNS/Consul |
| 8 |
MCP Transports |
✅ Stdio/SSE/WebSocket, capability negotiation |
| 9 |
A2A/ACP |
✅ Server/Client, streaming, agent cards |
| 10 |
Observability |
✅ Auto-instrumentation, 15 Prometheus metrics |
| 11 |
Multi-tenancy |
✅ TenantRegistry, quotas, RBAC, isolation |
| 12 |
AuthZ/AuthN |
✅ JWT/mTLS/Token Introspection, ASGI middleware |
| 13 |
Deployment |
✅ Docker, Helm, K8s Operator, systemd |
| 14 |
Secrets |
✅ Vault/AWS/GCP/Azure + Composite fallback |
| 15 |
Rate Limiting |
✅ Redis sliding-window, priority-weighted |
| 16 |
Audit/Compliance |
✅ AuditLogger, GDPR, DataLineage, OPA, SBOM |
| 17 |
Developer UX |
✅ alc CLI, Graph Visualizer (Mermaid/Graphviz/HTML) |
Quick Install
# Core
pip install alcyoneus
# Common stacks
pip install "alcyoneus[openai,google-genai,mcp,pg_checkpoint,qdrant]"
pip install "alcyoneus[realtime]" # Audio agents
pip install "alcyoneus[browser]" # Browser automation
pip install "alcyoneus[cli]" # CLI tools
export OPENAI_API_KEY=sk-... # or GEMINI_API_KEY, ANTHROPIC_API_KEY
30-Second Example
from alcyoneus.core import StateGraph, Agent, ToolNode, START, END
from alcyoneus.storage.checkpointer import InMemoryCheckpointer
from alcyoneus.prebuilt.tools import safe_calculator, google_web_search
graph = StateGraph()
graph.add_node("agent", Agent(model="gemini/gemini-2.5-flash", tool_node="tools"))
graph.add_node("tools", ToolNode([safe_calculator, google_web_search]))
graph.add_edge("agent", "tools")
graph.add_edge("tools", "agent")
graph.set_entry_point("agent")
compiled = graph.compile(checkpointer=InMemoryCheckpointer())
result = compiled.invoke(
{"messages": [{"role": "user", "content": "What's 25 * 4?"}]},
config={"thread_id": "demo"}
)
Documentation
Examples
# Basic agent
python examples/react/react_sync.py
# MCP integration
pip install "alcyoneus[mcp]"
python examples/react-mcp/server.py # terminal 1
python examples/react-mcp/react-mcp.py # terminal 2
# Streaming
python examples/react_stream/stream_react_agent.py
# Realtime audio
pip install "alcyoneus[realtime]"
export GEMINI_API_KEY=...
python examples/realtime/audio_agent_file.py
Testing & Quality
# Run all tests (3159 passing)
pytest tests/ -q
# Lint
ruff check . && ruff format .
# Type check
mypy alcyoneus/
License
MIT License - see LICENSE for details.
Links
Alcyoneus OS — Build intelligent agents with confidence.
