Create & Publish

Repository-Based Agents

Repository-based agents are standalone projects that can be published and shared with other Compozy users. You can create agents using either TypeScript (running on Deno runtime) or Python, which can be compiled into platform-specific binaries for efficient distribution.

Repository Structure

my-agent/
├── deno.json          # Deno configuration and tasks
├── manifest.yaml      # Agent metadata and schema definitions
├── README.md          # Documentation
├── src/               # Source code directory
│   ├── index.ts       # Main entry point
│   └── lib/           # Additional modules
├── tests/             # Test files
│   └── index.test.ts
├── dist/              # Compiled binaries
└── my-agent

my-agent/
├── deno.json          # Deno configuration and tasks
├── manifest.yaml      # Agent metadata and schema definitions
├── README.md          # Documentation
├── src/               # Source code directory
│   ├── index.ts       # Main entry point
│   └── lib/           # Additional modules
├── tests/             # Test files
│   └── index.test.ts
├── dist/              # Compiled binaries
└── my-agent

my-agent/
├── pyproject.toml    # Python project configuration
├── manifest.yaml     # Agent metadata and schema definitions
├── src/             # Source code directory
│   ├── main.py      # Main entry point
│   └── lib/         # Additional modules
├── tests/           # Test files
│   └── test_main.py
├── README.md        # Documentation
└── dist/            # Compiled binaries
    └── my-agent

Project Configuration

The deno.json file configures your development environment and build process:

{
  "tasks": {
    "dev": "deno run --allow-read --allow-net src/index.ts",
    "test": "deno test --allow-read --allow-net tests/",
    "compile": "deno compile --allow-read --allow-net --output dist/market-analyst src/index.ts",
  },
}

The compile task is required in your deno.json file. Compozy uses this task to compile your agent into a platform-specific binary during the publishing process.

The deno.json file configures your development environment and build process:

{
  "tasks": {
    "dev": "deno run --allow-read --allow-net src/index.ts",
    "test": "deno test --allow-read --allow-net tests/",
    "compile": "deno compile --allow-read --allow-net --output dist/market-analyst src/index.ts",
  },
}

The compile task is required in your deno.json file. Compozy uses this task to compile your agent into a platform-specific binary during the publishing process.

The pyproject.toml file configures your Python project:

[project]
name = "market-analyst"
version = "1.0.0"
description = "AI agent for market analysis and insights"
requires-python = ">=3.8"
dependencies = [
    "openai",
    "pydantic",
    "controlflow"
]

[project.optional-dependencies]
test = ["pytest", "pytest-asyncio"]

[tool.pytest.ini_options]
asyncio_mode = "auto"

Manifest File

The manifest.yaml defines your agent’s metadata, configuration, and action schemas. Here’s a comprehensive example:

The type field is required and must be set to agent. For Python projects, set main: src/main.py instead of src/index.ts.

type: agent
name: market-analyst
version: 1.0.0
license: MIT
description: AI agent for analyzing market trends and providing financial insights
repository: https://github.com/your-org/market-analyst
main: src/index.ts  # Use src/main.py for Python projects

tags:
  - finance
  - analysis
  - market

author:
  name: Your Name
  email: your.email@example.com
  organization: Your Organization

# Default agent configuration
config:
  provider: openrouter
  model: anthropic/claude-3-sonnet
  temperature: 0.7
  max_tokens: 2000
  top_p: 0.95
  frequency_penalty: 0
  presence_penalty: 0

schema:
  # Agent-specific parameters
  input:
    type: object
    properties:
      language:
        type: string
        description: "Response language"
        default: "en"
      threshold:
        type: number
        description: "Confidence threshold for recommendations"
        minimum: 0
        maximum: 1
        default: 0.7
    required: []

actions:
  analyze_market:
    id: analyze_market
    instructions: "Analyze market trends and provide detailed insights"
    messages:
      - role: system
        content: "You are an expert market analyst. Focus on technical analysis and market sentiment."
      - role: user
        content: |
          Analyze these stocks: {{ input.symbols }} for timeframe: {{ input.timeframe }}
    input:
      type: object
      properties:
        symbols:
          type: array
          items:
            type: string
          description: "List of stock symbols to analyze"
        timeframe:
          type: string
          enum: ["1d", "1w", "1m", "3m", "1y"]
          default: "1d"
      required: [symbols]
    output:
      type: object
      properties:
        sentiment:
          type: string
          enum: [positive, negative, neutral]
        confidence:
          type: number
          minimum: 0
          maximum: 1
        analysis:
          type: string
        recommendations:
          type: array
          items:
            type: string
      required: [sentiment, confidence, analysis]

Implementation

import { Agent } from "@mastra/core/agent";
import { AgentConfig, ActionFromSchema } from "@compozy/types";
import { createMessages } from "@compozy/utils";
import { z } from "zod";
import { Mastra } from "@mastra/core";

// Define types based on manifest schema
interface AgentParams {
  language?: string;
  threshold?: number;
}

interface MarketAnalysis {
  symbols: string[];
  timeframe?: '1d' | '1w' | '1m' | '3m' | '1y';
}

const actionSchema = z.object({
  name: z.literal("analyze_market"),
  instructions: z.string(),
  messages: z.array(z.object({
    role: z.string(),
    content: z.string()
  })),
  input: z.object({
    symbols: z.array(z.string()),
    timeframe: z.enum(['1d', '1w', '1m', '3m', '1y']),
  }),
  output: z.object({
    sentiment: z.enum(['positive', 'negative', 'neutral']),
    confidence: z.number(),
    analysis: z.string(),
    recommendations: z.array(z.string()).optional(),
  }),
  temperature: z.number().optional(),
  max_tokens: z.number().optional(),
});

type AgentAction = ActionFromSchema<typeof actionSchema>;

function getModel(config: AgentConfig<AgentParams>) {
  if (config.provider === "openai") {
    return openai(config.model);
  }
  if (config.provider === "anthropic") {
    return anthropic(config.model);
  }
  throw new Error(`Unsupported provider: {{ config.provider }}`);
}

export async function run(
  input: MarketAnalysis,
  action: AgentAction,
  config: AgentConfig<AgentParams>
) {
  const agent = new Agent({
    name: "Market Analyst",
    model: getModel(config),
  });

  if (action.name === "analyze_market") {
    const messages = createMessages({ action, input, config });

    const response = await agent.generate(
      messages,
      {
        output: action.output,
        temperature: action.temperature ?? config.temperature,
        maxTokens: action.max_tokens ?? config.max_tokens,
      }
    );

    return response.object;
  }

  throw new Error(`Unsupported action: {{ action.name }}`);
}

import { Agent } from "@mastra/core/agent";
import { AgentConfig, ActionFromSchema } from "@compozy/types";
import { createMessages } from "@compozy/utils";
import { z } from "zod";
import { Mastra } from "@mastra/core";

// Define types based on manifest schema
interface AgentParams {
  language?: string;
  threshold?: number;
}

interface MarketAnalysis {
  symbols: string[];
  timeframe?: '1d' | '1w' | '1m' | '3m' | '1y';
}

const actionSchema = z.object({
  name: z.literal("analyze_market"),
  instructions: z.string(),
  messages: z.array(z.object({
    role: z.string(),
    content: z.string()
  })),
  input: z.object({
    symbols: z.array(z.string()),
    timeframe: z.enum(['1d', '1w', '1m', '3m', '1y']),
  }),
  output: z.object({
    sentiment: z.enum(['positive', 'negative', 'neutral']),
    confidence: z.number(),
    analysis: z.string(),
    recommendations: z.array(z.string()).optional(),
  }),
  temperature: z.number().optional(),
  max_tokens: z.number().optional(),
});

type AgentAction = ActionFromSchema<typeof actionSchema>;

function getModel(config: AgentConfig<AgentParams>) {
  if (config.provider === "openai") {
    return openai(config.model);
  }
  if (config.provider === "anthropic") {
    return anthropic(config.model);
  }
  throw new Error(`Unsupported provider: {{ config.provider }}`);
}

export async function run(
  input: MarketAnalysis,
  action: AgentAction,
  config: AgentConfig<AgentParams>
) {
  const agent = new Agent({
    name: "Market Analyst",
    model: getModel(config),
  });

  if (action.name === "analyze_market") {
    const messages = createMessages({ action, input, config });

    const response = await agent.generate(
      messages,
      {
        output: action.output,
        temperature: action.temperature ?? config.temperature,
        maxTokens: action.max_tokens ?? config.max_tokens,
      }
    );

    return response.object;
  }

  throw new Error(`Unsupported action: {{ action.name }}`);
}

from typing import TypedDict, List, Optional, Literal, Dict, Any
from compozy.types import AgentConfig
from compozy.utils import create_messages
import controlflow as cf

class AgentParams(TypedDict):
    language: Optional[str]
    threshold: Optional[float]

class Input(TypedDict):
    symbols: List[str]
    timeframe: Optional[Literal['1d', '1w', '1m', '3m', '1y']]

class Message(TypedDict):
    role: str
    content: str

class Action(TypedDict):
    name: str
    instructions: str
    messages: List[Message]
    input: Dict[str, Any]
    output: Dict[str, Any]
    temperature: Optional[float]
    max_tokens: Optional[int]

class Output(TypedDict):
    sentiment: Literal['positive', 'negative', 'neutral']
    confidence: float
    analysis: str
    recommendations: Optional[List[str]]

async def run(input: Input, action: Action, config: AgentConfig[AgentParams]) -> Output:
    agent = cf.Agent(
        name="Market Analyst",
        provider=config.provider,
        model=config.model,
        temperature=action.get('temperature', config.temperature),
        max_tokens=action.get('max_tokens', config.max_tokens),
        top_p=config.top_p,
        frequency_penalty=config.frequency_penalty,
        presence_penalty=config.presence_penalty,
    )

    if action['name'] != 'analyze_market':
        raise ValueError(f"Unsupported action: {action['name']}")

    messages = create_messages({ action=action, input=input, config=config })

    result = await agent.run(
        messages=messages,
        output_schema=action['output']
    )

    if result['confidence'] < config.params.get('threshold', 0.7):
        result['recommendations'] = ['Confidence too low for specific recommendations']

    return result

Publishing Process

Prepare Your Repository

Initialize a GitHub repository
Set up CI/CD workflows for testing and releases
Include comprehensive documentation in README.md

Validate

Ensure your agent configuration is valid:

compozy agent validate

Version and Release

Use semantic versioning (MAJOR.MINOR.PATCH)
Create GitHub releases with detailed changelogs
Tag releases (e.g., v1.0.0)

Build and Distribution

# Compile for all platforms
deno task compile

Submit to Registry

Once your agent is ready for distribution:

# Login to Compozy CLI
compozy login

# Submit your agent
compozy publish

After submission, our team will review your agent to ensure it meets our quality and security standards. You’ll receive notifications about the review process and when your agent is approved and listed in the registry.

Make sure your agent follows these best practices:

Use semantic versioning for your agent versions
Include comprehensive documentation in README.md
Implement proper error handling and validation
Use environment variables for sensitive configuration
Define clear action schemas and instructions

Getting Started

Core Concepts

Create & Publish

Repository-Based Agents

Repository Structure

Project Configuration

Manifest File

Implementation

Publishing Process

Getting Started

Core Concepts

​Repository-Based Agents

​Repository Structure

​Project Configuration

​Manifest File

​Implementation

​Publishing Process

Repository-Based Agents

Repository Structure

Project Configuration

Manifest File

Implementation

Publishing Process