What is Effero?

Effero is an open-source, modular runtime for building agentic AI systems that operate in the physical and digital world — not just chatbots that call APIs, but agents that see, hear, speak, reason, and act: driving a robot arm, flipping a smart switch, narrating what a camera sees, piloting a drone, or operating a desktop on your behalf.

It is designed from the ground up around three observations from the current state of the field:

1. Agent frameworks (LangGraph, CrewAI, AutoGen, OpenAI/Google Agent SDKs) are excellent at software tool-calling but treat the physical world as an afterthought. There's no first-class notion of a camera frame, a LiDAR scan, a servo limit, or an emergency stop.
2. Robotics and IoT stacks (ROS 2, Home Assistant, Matter/Thread, MQTT) are mature at device control but were never built around LLM-native reasoning, memory, or natural-language planning.
3. Safety guardrails built for text ("don't say harmful things") do not transfer to embodied action ("don't do harmful things"). Research on LLM-controlled robots has repeatedly shown that a model that refuses a harmful sentence will still happily execute a harmful plan unless there's a runtime layer, independent of the model, checking actions against the physical world before they execute.

Effero's job is to sit in the middle: one agent runtime, one skill/tool abstraction, one safety layer — many bodies. The same agent that debugs your code over chat can, with the right skills installed, water your plants, greet a visitor at the door, or guide a manipulator arm through a pick-and-place task — through the same planner, the same memory, and the same guardrail engine.

Why Effero

Effero doesn't try to replace ROS 2, Home Assistant, or your favorite agent SDK — it speaks their protocols and gives you one coherent reasoning layer on top of all of them.

Core Principles

1. Protocol-native, not protocol-adjacent. MCP is how skills are exposed, and A2A is how agents talk to each other. These aren't plugins bolted onto a proprietary tool format — they are the tool format.
2. Everything is a Skill. A robot joint, a smart plug, a shell command, a browser click, and a REST API call are all exposed through the same Skill interface: a name, an input schema, a safety class, and an execution scope. The planner never needs to know how a skill is implemented.
3. Modality-agnostic perception. Speech, vision, and sensor telemetry are pluggable pipelines with a common event bus. Swap Whisper for Moonshine, or a cloud vision API for a local YOLO model, without touching agent logic.
4. Local-first, cloud-optional. Effero runs end-to-end on a Raspberry Pi / Jetson / mini-PC with local models (llama.cpp, Ollama, faster-whisper, Piper) and no internet connection. Cloud LLMs and APIs are opt-in accelerants, not requirements. See the Hardware Compatibility table for tested platforms.
5. Safety is architecture, not a system prompt. A runtime guardrail engine — independent of the LLM — grounds every proposed action against the robot/device's actual state and a declarative policy before it is allowed to execute.
6. Bring your own model. LLM, ASR, TTS, and vision backends are all adapters behind stable interfaces, matched to your hardware budget: from a 27 MB edge speech model to a frontier cloud LLM.
7. Polyglot by design, not by default. Everything defaults to Python — that's where the ecosystem and the contributors are. Rust shows up only where the architecture specifically calls for it: the safety kernel (an independently-auditable, LLM-free guardrail process) and constrained-device MCP servers (crates/) that a Python runtime can't run on. It's never a blanket rewrite of the core.

Architecture

Effero is organized in six layers. Data flows up from perception, through cognition, back down through skills into the physical/digital world — with the safety layer able to intercept at every boundary.

┌─────────────────────────────────────────────────────────────────────────────┐
│                             INTERFACES                                      │
│   CLI / SDK          Chat & Web Dashboard       Voice Loop       Mobile     │
└──────────┬──────────────────┬──────────────────────┬──────────────┬─────────┘
           │                  │                      │              │
           ▼                  ▼                      ▼              ▼
┌─────────────────────────────────────────────────────────────────────────────┐
│                          PERCEPTION BUS                                     │
│                                                                             │
│  ┌──────────────────┐  ┌──────────────────┐  ┌────────────────────────┐     │
│  │  Vision Pipeline │  │  Audio Pipeline  │  │     Sensor Fusion      │     │
│  │  detection · seg │  │  wake-word · VAD │  │  IMU · LiDAR · telem   │     │
│  │  VLA · OCR       │  │  ASR · TTS       │  │  odometry · env        │     │
│  └──────────────────┘  └──────────────────┘  └────────────────────────┘     │
└──────────────────────────────┬──────────────────────────────────────────────┘
                               │ events
                               ▼
┌─────────────────────────────────────────────────────────────────────────────┐
│                    COGNITION CORE  (Effero Runtime)                         │
│                                                                             │
│  ┌────────────────────┐  ┌──────────────┐  ┌────────────────────────────┐   │
│  │   Orchestrator /   │  │    Memory    │  │       Model Router         │   │
│  │      Planner       │  │              │  │                            │   │
│  │  plan → act →      │  │  working     │  │  local llama.cpp / Ollama  │   │
│  │  observe → replan  │  │  episodic    │  │         ⇅                  │   │
│  │                    │  │  semantic    │  │  cloud  OpenAI / Anthropic │   │
│  └────────────────────┘  └──────────────┘  │         / Google           │   │
│                                            └────────────────────────────┘   │
│                               ◄── A2A ──►                                   │
└──────────────────────────────┬──────────────────────────────────────────────┘
                               │
                               ▼
┌─────────────────────────────────────────────────────────────────────────────┐
│                      SKILL LAYER  (MCP-native)                              │
│                                                                             │
│  ┌──────────┐  ┌──────────┐  ┌──────────────┐  ┌────────────────────────┐   │
│  │ Robotics │  │   IoT    │  │ Computer Use │  │  Custom / Community    │   │
│  │  Skills  │  │  Skills  │  │    Skills    │  │       Skills           │   │
│  └──────────┘  └──────────┘  └──────────────┘  └────────────────────────┘   │
│                        ▲  Skill Registry  ▲                                 │
└────────────────────────┼──────────────────┼─────────────────────────────────┘
                         │                  │
      ┌──────────────────┘                  └──────────────────┐
      ▼                                                        ▼
┌───────────────────────────────────────┐  ┌──────────────────────────────────┐
│      DEVICE ABSTRACTION LAYER         │  │      SAFETY & GOVERNANCE         │
│                                       │  │                                  │
│  ROS 2 Adapter                        │  │  Runtime Guardrail Engine        │
│  MQTT / Matter / Zigbee / Thread      │◄─┤  (policy-as-code, Rust kernel)   │
│  Serial / GPIO / Arduino / ESP32      │  │                                  │
│  Generic REST / Cloud APIs            │  │  Human-in-the-loop Approval      │
│  OS Control (files · browser · input) │  │  Hardware E-Stop Hook            │
│                                       │  │  Audit Log                       │
└───────────────────────────────────────┘  └──────────────────────────────────┘

Layer breakdown

1. Interfaces — CLI (effero), a chat/web dashboard, a voice wake-word loop, and a thin mobile client. All talk to the same runtime over a local gRPC/WebSocket API.

2. Perception Bus — Independent, always-on pipelines that publish structured events (ObjectDetected, Utterance, SensorReading) onto a shared bus. The planner subscribes to what it needs instead of polling hardware directly.

3. Cognition Core — The actual "brain": a graph-based orchestrator (plan → act → observe → replan), a three-tier memory (working context, episodic transcript, semantic vector recall), and a model router that lets every LLM call declare its needs (latency, tool-calling reliability, multimodality) and get routed to a local or cloud backend accordingly.

4. Skill Layer — Every capability the agent can invoke — moving a servo, toggling a light, reading a file, clicking a browser element — is registered as an MCP tool with a declared safety class (see Safety). This is also the layer where the community contributes: a new skill is a new MCP server, full stop.

5. Device Abstraction Layer (DAL) — Where skills actually touch hardware: a ROS 2 bridge for robots, an MQTT/Matter/Zigbee/Thread bridge for IoT, a serial/GPIO bridge for microcontrollers (Arduino, ESP32), a generic REST adapter for cloud-connected appliances, and an OS-control adapter for desktop/browser automation.

6. Safety & Governance — A runtime guardrail engine that sits between the skill layer and the DAL, independent of the LLM, checking every proposed action against declarative policy and the device's live state before it's allowed to reach hardware.

Quickstart

Effero targets Python 3.11+ and runs on Linux, macOS, Windows (WSL2), and edge boards (Raspberry Pi 5, Jetson Orin).

# 1. Install
pip install effero

# 2. Scaffold a new project
effero init my-agent
cd my-agent

# 3. Point it at a model (local by default)
effero config set model.backend ollama --model qwen3:8b
# or, to use a cloud model instead:
# effero config set model.backend anthropic --model claude-sonnet-5

# 4. Run it
effero run

Minimal agent definition

# effero.yaml
agent:
  name: home-and-desk-assistant
  model:
    backend: ollama
    model: qwen3:8b
    fallback: [anthropic:claude-sonnet-5]

perception:
  audio:
    wake_word: "hey effero"
    asr: faster-whisper:small.en
    tts: piper:en_US-amy-medium
  vision:
    enabled: true
    backend: yolov9

skills:
  - iot.lights           # via MQTT/Matter
  - iot.thermostat
  - robotics.arm_pick_place   # via ROS 2
  - computer_use.browser

safety:
  policy: safety/policies/home.yaml
  require_approval_for: [robotics.*, computer_use.file_delete]

Defining a custom skill (Python SDK)

from effero.sdk import skill, SafetyClass

@skill(
    name="iot.thermostat.set_temperature",
    description="Set the target temperature of a named thermostat.",
    safety_class=SafetyClass.ACT_AUTONOMOUS,  # no approval needed
)
def set_temperature(thermostat_id: str, celsius: float) -> dict:
    device = mqtt_matter.get_device(thermostat_id)
    device.set_attribute("target_temperature", celsius)
    return {"status": "ok", "device": thermostat_id, "target_temperature": celsius}

Because this is registered as an MCP tool under the hood, it is immediately callable by Effero's own planner and by any other MCP-compatible client (Claude, an IDE agent, etc.) — no extra glue code.

Connecting a robot over ROS 2

from effero.adapters.ros2 import ROS2Bridge

bridge = ROS2Bridge(node_name="effero_bridge")
bridge.expose_action("/arm/pick_place", skill_name="robotics.arm_pick_place")
bridge.expose_topic("/scan", event_name="lidar.scan")  # feeds the perception bus
bridge.spin()

Perception: Pluggable Backends

Effero ships adapters for today's strongest open building blocks and lets you swap any of them per deployment:

Speech pipelines speak the Wyoming protocol, the same lightweight streaming standard used by Home Assistant's local voice stack, so Effero can plug directly into an existing Wyoming satellite/microphone setup instead of reinventing audio transport.

Skills & the Device Abstraction Layer

A Skill is Effero's only unit of capability. Every skill declares:

• an MCP tool schema (name, typed inputs/outputs, description the planner reasons over),
• a safety class — READ_ONLY, ACT_AUTONOMOUS, ACT_WITH_APPROVAL, or ACT_RESTRICTED (simulation-only until explicitly promoted),
• and an adapter binding — which piece of the Device Abstraction Layer actually carries it out.

Multiple robots or Effero instances coordinate over A2A (Agent-to-Agent protocol), so a fleet of agents — a mobile robot, a stationary arm, and a home-automation instance — can negotiate and delegate subtasks instead of each acting blindly.

Safety & Governance

Text-alignment guardrails do not protect against unsafe physical actions — a model that refuses to describe something harmful in words has been repeatedly shown to still execute an unsafe robot plan unless something outside the model stops it. Effero's guardrail engine is built around that finding:

• Policy-as-code: declarative rules ("never exceed 0.5 m/s within 1 m of a person", "never unlock an exterior door between 11pm–6am without approval") are grounded against the live world model at runtime, not just the prompt.
• Independent runtime check: every action from the skill layer is validated against policy and current device/robot state before it reaches the Device Abstraction Layer — this check does not go through the LLM and cannot be prompt-injected away.
• Human-in-the-loop approval: skills marked ACT_WITH_APPROVAL pause for explicit confirmation (CLI, chat, or push notification) before executing.
• Simulation-first promotion: new or community-contributed skills default to ACT_RESTRICTED (dry-run/simulated) until a maintainer or operator promotes them.
• Hardware e-stop hook: a dedicated, LLM-independent kill switch that adapters must implement — pulling it halts actuation regardless of what the agent is "thinking."
• Full audit log: every plan, tool call, guardrail decision, and approval is logged for post-hoc review.

This mirrors the direction of current robot-safety research: safety guarantees need to live at the level of grounded, verifiable actions — not at the level of text.

Status: the independent runtime check above is implemented as crates/effero-safety-kernel — a small, dependency-minimal Rust process (deliberately not embedded in the Python runtime, for the reason stated above) that loads a declarative policy and serves allow/require-approval/deny/limit decisions over a local Unix socket. It's built, unit-tested, and wired into src/effero/safety/ with full end-to-end client verification.

Example Builds (examples/)

• local-smart-home/ ✅ — voice-controlled home automation: wake-word → ASR → LLM → Matter/MQTT devices → TTS. Runs as an interactive REPL.
• desktop-copilot/ ✅ — a computer-use agent that operates shell, files, and browser on the user's behalf, gated behind ACT_WITH_APPROVAL for destructive actions.
• mobile-manipulator/ (planned) — natural-language task planning over a ROS 2-connected mobile robot with a VLA-driven manipulation skill for pick-and-place.
• industrial-monitor/ (planned) — sensor-fusion anomaly detection across a fleet of IoT sensors, with agentic triage and human-approved actuation.
• multi-robot-swarm/ (planned) — two or more Effero instances coordinating over A2A to split a shared task.

Contributing

Effero is community-built from day one. Good places to start:

• New skills are the easiest contribution — a skill is a small, testable MCP server. See sdk/ and CONTRIBUTING.md.
• New adapters (a new IoT protocol, a new robot middleware, a new microcontroller board) live in adapters/.
• New perception backends (an ASR/TTS/vision model you want supported) live in perception/.
• Please read CODE_OF_CONDUCT.md before opening issues or PRs. Look for issues tagged good-first-issue.

License

Released under the Apache License 2.0 (see LICENSE) — chosen deliberately over MIT for its explicit patent grant, which matters for a framework that touches physical hardware. Feel free to revisit this choice for your own fork if your governance needs differ.

Acknowledgments

Effero interoperates with, and draws design lessons from, a number of independent open ecosystems it is not affiliated with: the Model Context Protocol and A2A protocol for agent/tool interoperability, ROS 2 for robotics middleware, Home Assistant and the Wyoming protocol for local-first voice and smart-home patterns, the open Vision-Language-Action research line (OpenVLA, Octo, and related generalist robot policies), and the broader open ASR/TTS ecosystem (Whisper and its faster-whisper/Moonshine/Parakeet derivatives, Piper). Effero exists to connect these worlds, not replace them.