API Reference

Overview

This reference covers the configuration formats, template variables, package system, and provider capabilities available when building agents and templates on AgentOp.

Agent Configuration

When creating or editing an agent, you configure the following properties:

Property	Description
name	Display name for the agent (shown in marketplace and agent detail page)
description	Markdown-supported description of agent capabilities
template	The base template providing HTML structure, default code, and styling
provider	AI backend: openai, anthropic, or local
python_code	Custom Python agent logic that runs via Pyodide in the browser
tags	Categorization tags for discovery in the marketplace
is_public	Whether the agent is listed in the marketplace (default: private)
allow_forks	Whether other users can fork and customize this agent

Packages Configuration Format

Python packages for an agent are configured using this JSON structure:

{
  "provider": "openai",        // "openai", "anthropic", or "local"
  "pyodide_builtins": [
    "numpy",
    "pandas",
    "matplotlib"
  ],
  "pypi_packages": {
    "requests": ">=2.28.0",   // minimum version
    "langchain": ">=0.1.0",
    "beautifulsoup4": "*"     // latest available version
  }
}

Template Configuration

Templates define the reusable foundation for agents: HTML/CSS/JS structure, default Python code, prompt defaults, and package requirements.

Property	Description
name	Unique name for the template
description	What kinds of agents this template is designed for
html_code	HTML structure with Mustache placeholders like {{agent_name}}
css_code	Optional CSS styles injected into the generated agent
js_code	Optional JavaScript for client-side interactivity
python_code	Default Python code that agents built from this template start with
system_prompt	Default system-level LLM instructions
user_prompt_template	How user messages are formatted before being sent to the LLM
prompt_variables	JSON schema for configurable prompt placeholders
few_shot_examples	Example input/output pairs to guide LLM behaviour
conversation_memory_enabled	Whether to maintain conversation history across messages
max_memory_messages	Maximum number of past messages to retain in memory
default_packages	Default Python packages applied to all agents using this template

Prompt Variables Format

Define configurable placeholders in the prompt using this JSON structure:

{
  "domain": {
    "type": "string",
    "default": "general assistance",
    "description": "The agent's area of expertise"
  },
  "tone": {
    "type": "string",
    "default": "professional",
    "options": ["professional", "casual", "formal"]
  }
}

Few-Shot Examples Format

Guide LLM behaviour with example conversations:

[
  {
    "input": "What's the weather in San Francisco?",
    "output": "Let me check that for you. The weather in San Francisco is sunny and 72°F."
  },
  {
    "input": "How about tomorrow?",
    "output": "Tomorrow looks partly cloudy with a high of 68°F."
  }
]

Template Context Variables

These variables are injected into HTML templates when an agent is generated:

Variable	Type	Description
agent_name	string	Agent display name
description	string	Agent description
python_code	string	Agent Python code
provider	string	Provider identifier: openai, anthropic, or local
provider_display_name	string	Human-readable provider name (e.g., "OpenAI", "Anthropic Claude")
pyodide_version	string	Pyodide runtime version (e.g., "0.29.0")
pyodide_packages	list	Pyodide built-in package names selected for this agent
pypi_packages	dict	PyPI packages with their version specifiers
needs_webllm	boolean	Whether the WebLLM runtime should be loaded (true for local provider)
system_prompt	string	System prompt passed to the LLM
user_prompt_template	string	User prompt template

Package Management

AgentOp supports two types of Python packages that run in the browser via Pyodide:

Pyodide Built-in Packages

Pre-compiled packages bundled with the Pyodide distribution:

• Data Science: numpy, pandas, scipy, scikit-learn
• Visualization: matplotlib, bokeh, plotly
• Web: requests, beautifulsoup4, lxml
• Utilities: pillow, regex, yaml, pytz

PyPI Packages

Pure-Python packages installed at runtime from PyPI via micropip:

{
  "pypi_packages": {
    "requests": ">=2.28.0",  // minimum version
    "langchain": ">=0.1.0",
    "python-dotenv": "*"     // latest version
  }
}

Version Resolution

When an agent's custom packages conflict with its template's default packages, AgentOp automatically resolves the versions. Compatible version ranges are merged (the more restrictive constraint wins). Incompatible conflicts are reported in the package editor so you can resolve them manually.

HTML Generation

When you download an agent, AgentOp generates a self-contained HTML file on the fly. Nothing is stored on disk — the file is created fresh each time for privacy.

Generation Flow

1. Load Template: Fetch the HTML, CSS, and JS from the agent's template
2. Build Context: Collect agent data, packages, and provider config
3. Render Placeholders: Replace template variables with agent-specific values
4. Inject Runtime: Add encryption UI (OpenAI/Anthropic) or WebLLM loader (local)
5. Return File: Serve the complete standalone HTML file for download

Provider Capabilities

Provider	Display Name	Streaming	Tool Calling	API Key Required
openai	OpenAI	Yes	Full support via OpenAI API	Yes
anthropic	Anthropic Claude	Yes	Full support via Anthropic API	Yes
local	Local WebLLM	Yes	Full support via LangChain.js	No

Prompt Version Control

Every template supports prompt versioning so you can track changes and roll back if needed.

Creating a Version

1. Open a template and navigate to the Prompts tab
2. Edit the system prompt, user prompt template, or few-shot examples
3. Click Save as New Version
4. Add an optional description of what changed

Version Fields

Field	Description
version_number	Auto-incrementing version number (1, 2, 3...)
system_prompt	Snapshot of the system prompt at the time of saving
user_prompt_template	Snapshot of the user prompt template
few_shot_examples	Snapshot of few-shot examples
change_description	Optional note describing what changed
is_active	Whether this version's prompts are currently applied to the template

Activating a Version

Navigate to the version history page for a template, find the version you want, and click Activate. This applies that version's prompts to the template and marks it as active.

Security

API Key Encryption

AgentOp uses AES-256-GCM encryption with PBKDF2 key derivation (100,000 iterations) to protect API keys. Encryption and decryption happen entirely in the browser — no keys are ever transmitted to or stored on AgentOp servers.

1. When downloading an API-based agent, you choose an encryption password
2. Your API key is encrypted client-side before the download is generated
3. The encrypted key is embedded in the HTML file
4. When anyone opens the agent, they enter the password to decrypt the key locally

Privacy by Design

Agent HTML files are generated on demand and never saved to disk on the server. Each download is a fresh render. This means your agent code is never stored in a way that could be accessed outside of your account.

Next Steps