Documentation

Quick Start

Core Concepts

Library

The Library is your personal collection of reusable building blocks. Everything you create here can be used across all your projects and agents.

Library items can be private (only you can use them) or public (published to the Marketplace for others to install).

Agent Templates

Agent Templates are reusable agent configurations. They define the agent's role, system prompt, and reference the tools, actions, and schemas from your library.

{
  "name": "Graphic Designer",
  "role": "Visual Designer",
  "description": "Creates visual assets including logos and graphics",
  "category": "design",
  "icon": "🖼️",
  
  "system_prompt": "You are a Visual Designer agent...",
  
  "guidelines": [
    "Follow brand color palette",
    "Use vector formats for logos and icons",
    "Maintain visual consistency across assets"
  ],
  
  "qa_pairs": [
    {
      "question": "What are the social media image sizes?",
      "answer": "Twitter: 1200x675px. Instagram: 1080x1080px.",
      "category": null
    }
  ],
  
  "tool_ids": [],
  "action_ids": [],
  "schema_ids": [
    { "id": "schema-uuid", "source": "library" }
  ],
  
  "publish_to_marketplace": false
}

Fields

Tools

Tools give agents capabilities to interact with external systems. Define installation commands and required environment variables.

Tool Configuration

{
  "name": "supabase-cli",
  "description": "Supabase CLI for managing databases and migrations",
  "installation_command": "npm install -g supabase",
  "setup_command": "supabase login --token ${SUPABASE_ACCESS_TOKEN}",
  "required_env_vars": ["SUPABASE_ACCESS_TOKEN"],
  "publish_to_marketplace": true
}

Fields

Actions

Actions are custom scripts that agents can execute. Write your own code in Python, Shell, or JavaScript.

{
  "name": "sync_agent_config",
  "description": "Sync configuration across agents in a project",
  "language": "python",
  "code": "<your Python code here>",
  "tool_dependencies": [],
  "action_dependencies": ["httpx"],
  "required_env_vars": ["API_SERVICE_KEY", "API_BASE_URL"],
  "publish_to_marketplace": true
}

Fields

Dependencies are automatically installed when the agent container is built.

Message Schemas

Message Schemas define the structure of messages agents send. They provide type safety and validation for inter-agent communication.

{
  "display_name": "Deployment Request",
  "description": "Request to deploy code changes to an environment",
  "schema_fields": [
    {
      "key": "branch",
      "description": "Git branch name",
      "type": "string",
      "required": true
    },
    {
      "key": "environment",
      "description": "Target environment (staging, production)",
      "type": "string",
      "required": true
    },
    {
      "key": "services",
      "description": "List of services to deploy",
      "type": "array",
      "required": true
    },
    {
      "key": "commit_sha",
      "description": "Git commit SHA to deploy",
      "type": "string",
      "required": true
    },
    {
      "key": "approval_required",
      "description": "Whether human approval is required",
      "type": "boolean",
      "required": false
    }
  ],
  "publish_to_marketplace": true
}

Fields

Field types: string,number,boolean,object,array

Marketplace

The Marketplace is a public library of tools, actions, schemas, and templates shared by the community. Install items to your library with one click.

When you edit an installed marketplace item, your changes are tracked in user_version. You can restore to the original marketplace version at any time.

Projects & Agents Overview

Projects are workspaces that contain your agents. Agents are AI workers built from templates in your library. Together, they form your AI team.

Projects

Projects are workspaces that contain agents. They provide:

• LLM configuration (provider, model, API keys)
• Repository connections (GitHub, GitLab)
• Project-level tool/action/schema overrides
• Shared Q&A knowledge base
• Message routing between agents

Agents

Agents are AI workers with specific roles. Create them from templates or configure from scratch.

Agent Configuration

{
  "name": "Mona",
  "description": "Coordinates work across agents",
  "template_id": "template-uuid",
  "avatar": { "face": {...}, "eyes": {...}, ... },
  "status": "active",
  
  "knowledge_base": {
    "system_prompt": "You are a Technical PM agent...",
    "guidelines": ["Break large features into tasks", ...],
    "qa_pairs": [{ "question": "...", "answer": "..." }]
  },
  
  "tools": ["tool-uuid-1", "tool-uuid-2"],
  "actions": ["action-uuid-1"],
  "message_schemas": ["schema-uuid-1", "schema-uuid-2"],
  
  "llm": {
    "provider": "openai",
    "model": "gpt-4o"
  },
  
  "runtime": {
    "preset": "nimble",
    "cpu_limit": 0.5,
    "memory_mb": 512
  }
}

Agent configuration is assembled from: template → project settings → agent-specific overrides. The Agent Packager bundles everything into a deployable package.

Message Bus

The Message Bus is the central communication backbone. All messages between agents and users flow through it, and it manages the agent jobs queue.

Job States

Jobs track execution from creation through completion, including timing, output, and any errors encountered.

Message Router

The Message Router maps message schemas to agents and users. When a message using a schema is sent, it automatically routes to all configured recipients.

{
  "message_schema_id": "schema-uuid",
  "agent_ids": [
    "agent-uuid-1",
    "agent-uuid-2"
  ],
  "recipient_user_ids": [
    "user-uuid-1"
  ]
}

How It Works

1. Define a Message Schema in your library (e.g., "code_review_request")
2. Create a Message Route that maps that schema to target agents/users
3. When any agent or user sends a message using that schema, the router kicks in
4. Jobs are created for each target agent, notifications for each target user

Architecture

Agent Runner

Each agent container runs its own job processor that polls for and executes work. The runner lives inside the container alongside the LLM.

How It Works

1. Container starts and begins polling agent_jobs table
2. Filters for jobs assigned to this specific agent
3. Picks up pending jobs and marks them as running
4. Processes job using the LLM (Agent Brain) directly in the container
5. Updates job status to completed/failed when done
6. Continues polling for next job

Agent Chat

Users can communicate directly with agents through the chat interface. This provides a real-time, conversational way to interact with your AI team.

Cloud Run Deployment

Each deployed agent runs as a Cloud Run service with its packaged configuration.