Build, customize, and interact with AI agents.
A modular AI agent platform with a core engine that handles AI orchestration — from tool execution to policy enforcement to workflow management. Packed with capabilities like Agents, Skills, Tools and Prompts that can be composed to build any AI assistant.
Platform Architecture
Three layers working together: the browser interface you interact with, the bridge that translates commands, and the core engine that handles all the AI logic.
zod-client
The browser interface built with SolidJS and Tailwind. Provides real-time chat, model selection, event debugging, and plugin loading — all running entirely in your browser.
harness.js Bridge
Acts as a translator between the browser and the AI engine. Takes commands from the user interface, passes them to the core engine, and brings responses back for display.
zod-harness Engine
The brain of the operation. Manages conversation history, decides when to use tools, connects to AI providers like OpenRouter, enforces safety rules, and orchestrates complex workflows.
Design Principles
Complete by Default
All essential functionality is built into the core engine — no extra pieces needed for basic use.
Optional Extensions
Add specialized capabilities as modules when needed, but they're never required.
Event-Driven Communication
Components communicate through a messaging system, making the system flexible and easy to extend.
Lifecycle Awareness
Everything tracks its state properly, from initialization through active use to clean shutdown.
The Core Engine
At the heart of Zod is the AgentHarness — the central orchestrator that coordinates extensions, tools, events, and lifecycle management.
The Agent Loop
User Message
You type a message in the browser interface.
Context Building
System gathers system prompt, conversation history, and available tools.
AI Processing
Everything is sent to the AI model for processing.
Response Generation
AI streams back responses in real-time, potentially requesting tool usage.
Tool Execution
Tools run with proper error handling — some in parallel, some sequentially.
Completion
Results feed back to the AI until it finishes responding.
Built-in Extensions
Agent Extension
Manages the AI agent — streaming responses, tool calls, and conversation steering.
OpenRouter Extension
Connects to AI model providers with streaming, reasoning tokens, and multi-model support.
Policy Extension
Enforces budget limits, step limits, approval gating, and loop detection.
Project Extension
Handles workflow orchestration — organizing complex tasks into graphs that can branch and loop.
Storage Extension
Manages persistent state — key-value memory, message queues, session persistence, and auto-save.
Plugin System
Loads additional capabilities from the internet at runtime without modifying core code.
Capabilities
Four types of pluggable capabilities that extend what the AI can do — each serving a distinct purpose.
Agents
Sub-AIs with their own personality, knowledge, and tools. Like hiring a specialist for specific tasks.
- Code Reviewer — reviews code for bugs and best practices
- Technical Writer — converts descriptions into documentation
- Data Analyst — understands datasets and explains findings
- Security Auditor — scans for vulnerabilities
- DevOps Engineer — manages deployment and infrastructure
Skills
Expert personas that define how the AI thinks and communicates — not what tools it can use.
- JavaScript Expert — deep knowledge of patterns and pitfalls
- Python Data Scientist — thinks statistically
- Customer Support — specific tone and escalation rules
- Security Mindset — always considers security implications
- Concise Communicator — brief, to-the-point answers
Prompts
Reusable templates with placeholders — patterns for generating consistent responses.
- Bug Report Generator — formatted reports with reproduction steps
- Meeting Notes Summarizer — structured action items
- Code Explainer — plain language explanations
- Email Response Generator — professional drafts
- Test Case Creator — unit test generation
Tools
Actions the AI can take in the world — the most powerful extension type for external system interaction.
- File Operations — read, write, search, list directories
- Shell Commands — run terminal commands and scripts
- Web Operations — fetch pages, call APIs, send notifications
- Database Operations — query, migrate, manage data
- Integrations — GitHub, Slack, Email, Cloud providers
How They Work Together
Code Review System
Security skill + Code reviewer agent + Review prompt template + File reading tools
Customer Support
Support skill + Response prompt + Database lookup tools + Email tools
Data Pipeline
Data analyst agent + SQL tools + Python tools + Visualization prompts
Safety & Policy
Multiple layers of built-in protection keep your AI usage safe and predictable.
Budget Limits
Maximum tokens and cost per task prevents runaway usage. The system stops automatically when limits are reached.
Step Limits
Maximum steps and tool calls per task prevents infinite loops and excessive tool usage.
Loop Detection
Recognizes when the AI is repeating the same action and suggests taking a different approach.
Approval Gating
Sensitive tools require explicit human approval before running. Users authorize dangerous operations.
Tool Permissions
Tools can declare their safety characteristics — whether they modify data or just read, whether they're safe to run in parallel, whether they could cause damage, and custom permission checks.
Real-World Use Cases
Concrete examples of what organizations have built with Zod.
| Use Case | Components Used |
|---|---|
| Automated code review | Agent + Skill + File tools |
| Internal knowledge base Q&A | Agent + Search tools + Prompt templates |
| DevOps automation dashboard | Agent + Shell tools + GitHub integration |
| Customer support chatbot | Skill + Agent + Database tools + Email tools |
| Data processing pipeline | Agent + SQL tools + Python tools + Visualization prompts |
| Documentation generator | Agent + File tools + Markdown prompts |
| Security scanning | Agent + Shell tools + Custom security prompts |
| Meeting summarizer | Agent + Audio transcription + Prompt templates |
Looking Forward
The platform is designed to grow. Here's what's coming next.
Enhanced MCP Integration
Better support for Model Context Protocol servers
Multi-Agent Orchestration
Multiple AI agents working together on complex tasks
Distributed Deployment
Running across multiple machines for scale
Advanced Policies
More sophisticated rule engines for enterprise needs
Analytics Dashboard
Usage metrics and insights for optimization
Ready to build with AI agents?
Start chatting with the demo or explore the documentation to learn more.