跳转至

For Developers

Ready to use production-tested AI capabilities in your projects? You're in the right place.

⚡ Quick Start

1. Pick Your Format

All capabilities are packaged in universal formats — choose what fits your stack:

Format Best For Install Command
MCP Server Claude Desktop, Code npx @goldentellus/<capability>
Claude Skill Claude CLI projects Copy manifest to .claude/skills/
OpenAPI Any HTTP client Import spec from capabilities.mdtools/<name>/openapi.yaml
Standalone Script in any language Clone repo, run locally

2. Browse the Catalog

Visit goldentellus-capabilities to see what's available.

Each capability includes: - README with usage examples - Trust block declaring scopes, egress, secrets - At least 1 universal format (usually multiple) - Production validation evidence

3. Install & Use

Example: using an MCP-packaged tool:

# Add to your Claude Code MCP config
npx @goldentellus/crm-lookup

Example: using a standalone skill:

# Clone the repo
git clone https://github.com/GoldenTellus/goldentellus-capabilities
cd goldentellus-capabilities.mdskills/document-summary

# Copy to your Claude project
cp manifest.yaml ~/.claude/skills/goldentellus-doc-summary.yaml

🧩 Capability Types

Tools (Connectors)

Connect to external systems: CRM, database, file storage, APIs.

  • Execution: Runs in sandbox (gVisor, Firecracker, or your runtime)
  • Trust: Requires explicit scopes, egress whitelist, and secret refs
  • Format: OpenAPI (preferred), MCP Server, standalone script

Skills (Orchestration)

Prompt engineering + RAG for specific tasks.

  • Execution: RAG Adapter (retrieval → rerank → LLM)
  • Trust: Read-only scopes, no egress
  • Format: Claude Skill manifest, YAML prompt template

Knowledge Packs

Domain-specific knowledge for retrieval-augmented generation.

  • Execution: Vector indexing (Qdrant, pgvector, etc.)
  • Trust: Read-only, no secrets
  • Format: Structured documents + schema

External Roles

Customer-facing bot configurations (lead capture, support, partner portal).

  • Execution: Gateway Bot runtime
  • Trust: Public context only, outbound filtering
  • Format: YAML role config

🔐 Trust Model

Every capability declares a trust block:

trust:
  execution_class: executable  # declarative | executable | composite
  required_scopes: ["context:read", "api:salesforce"]
  egress: ["api.salesforce.com", "*.hubspot.com"]
  secret_refs: ["salesforce_api_key"]
  cost_profile:
    inference_cost: low
    task_tier: non_frontier

This tells you: - Scopes: What permissions it needs - Egress: Which external domains it calls - Secrets: What credentials you must provide - Cost: Inference expense estimate

🧪 Testing Locally

Before using in production:

  1. Read the trust block — know what it accesses
  2. Run in sandbox — use gVisor, Docker, or your preferred isolation
  3. Test with fake data — never use production secrets for first tests
  4. Check error handling — does it fail gracefully?

🤝 Contribute a Capability

Have a production-tested capability? Share it!

Requirements: - Must be deployed in production (not a demo) - Must support at least 1 universal format - Must include trust block and README - Must pass quality checklist

See Contributing Guide for details.

📚 Learn More

💬 Get Help

Next Steps: 1. Browse the capabilities catalog 2. Pick a capability and install it 3. Read its trust block 4. Test with sandbox + fake data 5. Integrate into your project