Creating Marketplace Agents

What's in a template

Deployment flow

When a user deploys your template, Tropic writes all workspace files to ~/.openclaw/workspace/{slug}/ on their VM, installs skills via clawhub, configures tool connections, registers the agent in openclaw.json, and restarts the gateway. The agent is ready to chat within seconds.

OAuth connections

OAuth connections provide managed authentication through Tropic's OAuth flow. The user connects their account via a consent screen — Tropic handles token refresh and credential rotation.

Manual credentials

For services without OAuth support, users provide values directly. These are encrypted with AES-256-GCM at rest and written to ~/.openclaw/.env on the VM.

Defining prerequisites

Declare prerequisites in your bundle configuration. Each entry has a key (unique identifier), a type (drives the UI widget), and metadata for the deploy drawer.

[
  {
    "key": "SUPABASE",
    "label": "Supabase Project",
    "type": "supabase",
    "required": true,
    "helpText": "Connect your Supabase project for database access"
  },
  {
    "key": "GOOGLE",
    "label": "Google Account",
    "type": "google",
    "required": true,
    "helpText": "Connect Google for Gmail access"
  },
  {
    "key": "OPENAI_API_KEY",
    "label": "OpenAI API Key",
    "type": "secret",
    "required": false,
    "helpText": "Optional: for embedding generation",
    "placeholder": "sk-..."
  }
]

How credentials reach the agent

All credential values end up in ~/.openclaw/.env on the VM. The gateway restarts after .env is written so values are immediately available.

# From Supabase OAuth connection
SUPABASE_URL=https://abc123.supabase.co
SUPABASE_ANON_KEY=eyJ...
SUPABASE_AGENT_EMAIL=tropic-agent-uuid@tropic.bot
SUPABASE_AGENT_PASSWORD=generated-password

# From manual secret prerequisite
OPENAI_API_KEY=sk-abc123

# Google is configured via gog CLI auth, not env vars

Declaring skills

List the clawhub package slugs your agent needs in the agent.skills array.

"agent": {
  "model": "anthropic/claude-sonnet-4-6",
  "skills": ["agent-browser", "gog"]
}

Skills with env vars

If a skill requires its own environment variables (e.g. an API key for a third-party service), declare them in the skill's frontmatter. Users configure these per-agent after installing the skill.

---
name: youtube-full
description: Complete YouTube toolkit
metadata:
  openclaw:
    requires:
      env:
        - TRANSCRIPT_API_KEY
---

Policy format

Policies use three sections. Each section contains bullet-pointed rules that the agent must follow.

## ALLOW
- Search the web for public information
- Read and summarize emails
- Query the database for financial records

## REQUIRE CONFIRM
- Send emails on behalf of the user
- Delete database records
- Modify account settings

## DENY
- Share credentials or API keys
- Access files outside the workspace
- Make purchases or financial commitments

How enforcement works

Policies are enforced at two levels. The policy text is appended to AGENTS.md so the agent sees the rules in its context. Additionally, the policy is written to policy.md where a message-level hook classifies every incoming message against the rules and blocks denied requests before the agent sees them.

Full example

This is what a complete agent template looks like. The workspace object maps filenames to their content as strings.

{
  "agent": {
    "model": "anthropic/claude-sonnet-4-6",
    "skills": ["gog"]
  },
  "workspace": {
    "AGENTS.md": "# Bookkeeping Agent\n\n## Purpose\n...",
    "SOUL.md": "# Soul\n\nYou are precise...",
    "IDENTITY.md": "# Identity\n\n- Name: BookBot\n...",
    "TOOLS.md": "# Tools\n\n## Google (gog)\n...",
    "USER.md": "# User\n\nSmall business owner...",
    "SCHEMA.md": "# Database Schema\n\n## transactions\n..."
  },
  "prerequisites": [
    {
      "key": "SUPABASE",
      "label": "Supabase Project",
      "type": "supabase",
      "required": true,
      "helpText": "Connect your Supabase project"
    },
    {
      "key": "GOOGLE",
      "label": "Google Account",
      "type": "google",
      "required": true,
      "helpText": "Connect Google for email access"
    }
  ]
}

Template metadata

In addition to the bundle config, each template has top-level metadata that's shown in the marketplace.

What gets reviewed

• • Workspace files are well-structured and complete
• • Prerequisites are correctly typed with helpful descriptions
• • Skills don't contain security risks or malicious patterns
• • Default policy provides reasonable guardrails
• • Agent behavior matches the description

Security requirements

All agents run inside Tropic's security stack. Your template benefits from these protections automatically:

• • Sondera — Cedar policy evaluation on every tool call
• • SecureClaw — VM hardening and continuous auditing
• • Isolation — each user gets a dedicated EC2 instance
• • Encryption — all credentials encrypted with AES-256-GCM at rest
• • Redaction — secrets are blocked before execution and redacted from output