AgentYield for Openclaw

AgentYield is a high-performance observability layer that detects waste and optimizes costs for LLM and tool calls without compromising data privacy.

philg
v1.0.3
Apr 21, 2026
1
794
0

Install & Download

1. ClawHub CLI

The fastest way to install a skill directly from the registry.

npx clawhub@latest install agentyield

2. Manual Installation

Copy the skill folder to one of these locations

Global
~/.openclaw/skills/
Workspace
<project>/skills/

Priority: Workspace > Local > Bundled

3. Prompt Installation

Copy this prompt to OpenClaw to install it automatically.

Help me install agentyield using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).

Prefer to download?

Get the raw skill files in a ZIP archive.

What is AgentYield?

AgentYield acts as a non-intrusive instrumentation layer designed specifically for Openclaw Skills sessions. It monitors every LLM and tool call made by a host agent to identify inefficiencies, redundant calls, and unnecessary costs. By focusing on structured telemetry rather than raw content, it provides actionable insights into agent performance while maintaining a zero-trust posture toward sensitive data.

The skill operates in the background, ensuring it never slows down or alters the host agent's primary workflow. It is an essential tool for developers looking to scale AI agents sustainably, providing a clear Waste Score and detailed run analytics through the Openclaw Skills ecosystem.

AgentYield Use Cases

  • Identifying redundant tool calls that increase latency and cost.
  • Monitoring LLM token usage and context window efficiency across long sessions.
  • Tracking real-time USD costs for multi-model agentic workflows.
  • Auditing agent behavior for reasoning loops or inefficient planning stages.
  • Evaluating prompt effectiveness through structured purpose tagging like reasoning or self-check.

How AgentYield Works

  1. The skill initializes by reading the AGENTYIELD_API_KEY and generating a local-only cryptographic salt for data privacy.
  2. It hooks into the Openclaw Skills session lifecycle to intercept LLM and tool call events as they occur.
  3. Every tool input is salted and hashed using SHA-256 locally, ensuring original data is never transmitted.
  4. Events are collected in an in-memory buffer until a checkpoint threshold (event count or time interval) is reached.
  5. The skill performs a non-blocking POST request to the AgentYield API with redacted telemetry.
  6. Developers receive a Waste Score and a link to a visual dashboard for detailed run analysis.

AgentYield Setup

To integrate this with your Openclaw Skills environment, set the required environment variable:

export AGENTYIELD_API_KEY=ay_live_your_key_here

The skill will automatically initialize upon the next session. No additional configuration files are required, ensuring a secure and simple deployment.

AgentYield Data Schema & Taxonomy

AgentYield utilizes a strict telemetry schema that prioritizes privacy by excluding raw inputs and outputs.

Event Type Data Collected Privacy Mechanism
LLM Call Model ID, Token counts (In/Out), USD Cost, Purpose Tag No raw prompt/completion text
Tool Call Tool name, USD Cost, Input Hash Salted SHA-256 (256-bit) hashing
Session Run ID, Agent ID (openclaw), Checkpoint Label Unique UUIDs for session tracking

All hashing is performed locally using a hashSalt stored in the skill's private state directory: ~/.openclaw/skills/agentyield/state.json.

AgentYield Advanced Features

  • Automatic local salt generation for per-tenant rainbow table protection.
  • Idempotent checkpoint flushing with unique window labels to prevent data duplication.
  • Test-mode support using ay_test_ keys for evaluation without data persistence.
  • Bounded-await logic with a 3-second hard timeout to prevent agent latency.
  • Explicit session-end handling with a final flush signal for accurate run completion reporting.

SKILL.md


Loading

Related Openclaw Skills

METADATA

Github Stars: 0
forks: 0

Featured*