An open protocol standardizing how AI Agents perceive the real world.
World2Agent (W2A) is an open perception protocol designed specifically for AI Agents, built on the philosophy that "MCP gives hands — W2A gives ears." By defining a unified W2ASignal data envelope, the project standardizes events from various data sources (news, finance, GitHub, weather, etc.), enabling Agents to consume structured signals directly without adapting to underlying Sensor differences.
Protocol & Data Model#
- Unified Signal Contract: All Sensors follow the same
W2ASignalSchema, allowing Agents to consume signals without adapting to different data source variations. - High-Fidelity Envelope Structure: Includes
signal_id(UUID for deduplication & tracing), dual timestamps (emitted_at/occurred_at, UTC ms), three-segment event naming (domain.entity.action, e.g.,repo.issue.opened), natural languagesummary, self-describing raw data (source_eventwith schema + data), and flexible attachments (attachments, inline or reference-based). - Decentralized Extension Mechanism:
source_type,event.type, andmime_typeare all OPEN sets requiring no central registry; the_metafield supports experimental extensions, and consumers must ignore unknown keys. - Protocol Versioning: Current version is
w2a/0.1; directory names correspond toPROTOCOL_VERSION; breaking changes bump the directory, incremental changes modify in place.
Architecture#
Core data flow: World → Sensor → Agent
- Protocol Layer: Core is
schema/0.1/schema.ts, using self-contained TypeScript (no runtime dependencies) to define the wire format, withschema.jsongenerated viatypescript-json-schema. - Transport Layer: The protocol is deliberately transport-agnostic, not defining specific transport frames — each runtime implementation wraps its own (stdio, HTTP POST, WebSocket, etc.).
- Sensor Layer: Standard npm packages that monitor data sources via polling or webhooks, outputting unified
W2ASignal. - Bridge Layer: Runtime adapters (e.g.,
hermes-sensor-bridge,openclaw-sensor-bridge) that bridge Sensor signals to specific Agent frameworks. - Graph Layer (Roadmap): Planned support for composing and enriching multi-sensor signals before delivery to Agents.
Sensor Ecosystem & Distribution#
- Sensors as npm Packages: Each sensor is a standard npm package discoverable via
npm search w2a-sensor, installable with a single command. - Official SensorHub Directory: Organized across 9 categories (messaging / calendar / finance / research / social / dev / infra / data / custom), showcasing signal Schemas and one-click install commands.
- Meta-Sensor Support: Supports meta-sensors that monitor new W2A Sensor releases on npm (e.g.,
@quill-io/sensor-w2a-registry). - AI-Assisted Toolchain: The
build-w2a-sensorskill guides AI coding Agents through discovery → signal design → scaffolding → install recipe, publishing a new Sensor in ~50 lines of code.
Runtime Compatibility#
- Claude Code: Installed via Plugin mechanism (
/plugin marketplace add+/world2agent:sensor-add). - Hermes: Via
@world2agent/hermes-sensor-bridgenpm package + Skill mechanism. - OpenClaw: Via
@world2agent/openclaw-sensor-bridgenpm package + Skill mechanism. - Custom Agents: SDK developer quick-start path provided.
Typical Use Cases#
- Information Monitoring: Hacker News trending posts, frontier AI lab blog auto-push (
@world2agent/sensor-hackernews,@quill-io/sensor-frontier-ai-news). - Financial Alerting: Polymarket prediction market odds threshold triggers (
@quill-io/sensor-polymarket), FX pair target price alerts (@juliapupu/sensor-fx-frankfurter). - Developer Tools: GitHub repo Star growth monitoring, milestone tracking (
@quill-io/sensor-github-stars). - Lifestyle Info: Global city weather forecasts + astronomical moments (
@juliapupu/sensor-weather-everywhere). - Perception-Driven Automation: Bug detection → Agent auto-creates fix PRs; Steam free games/sales → Agent auto-claims.
Security Model#
Sensor signals directly drive Agent perception and action, so the security model relies on "only install trusted Sensors" — untrusted Sensors are equivalent to untrusted instruction sources. Hermes / OpenClaw require enabling the webhook platform and restarting the gateway on first Sensor Bridge installation. OpenClaw supports auto-pushing Agent responses to Feishu / iMessage / Telegram and other chat platforms.