- The Model Context Protocol (MCP) allows AI models like Claude to connect to external systems.
- MCP servers expose three core building blocks (primitives): Tools, Resources, and Prompts.
- Tools are executable actions (Capabilities).
- Resources provide contextual, read-only data (Information).
- Prompts are reusable message templates (Instructions).
- Easy memory trick: Tool = DO, Resource = KNOW, Prompt = GUIDE.
- A single enterprise workflow (like updating Salesforce) can seamlessly combine all three primitives.
🗺️ Module Map: Claude Foundation
CLAUDE FOUNDATION ├── What Is Claude? ├── Claude Model Families ├── Claude Reasoning ├── Claude Tool Use ├── AI Agents ├── Model Context Protocol (MCP) ├── MCP Architecture ├── MCP Security & Governance ├── Claude + MCP + Salesforce ├── Tools, Resources & Prompts 📍 (You are here) └── Claudeforce Architecture
Think of Claude as a highly capable worker sitting at a desk.
• A Tool is something the worker can DO (like typing an email or creating a task in Salesforce).
• A Resource is information the worker can KNOW (like looking at a printed customer file on the desk).
• A Prompt is the instruction manual that GUIDES the work (like a checklist for how to write a standard review).
Let's say you ask Claude: "Prepare me for tomorrow's meeting with ABC Healthcare."
- The Resource: Provides the background context (e.g., company size, industry).
- The Tool: Reaches into Salesforce to retrieve the latest open opportunities or recent activities.
- The Prompt: Forces Claude to format the final output into your company's standard "Meeting Briefing" layout.
🎬 Real-Life Scenario: The AI Sales Analyst
1. Resources supply the historical customer context.
2. Tools pull the live opportunity data or execute a command to log a new meeting prep activity.
3. Prompts structure the AI's thoughts into an executive summary.
4. Claude applies logic to highlight the most critical risks and up-sell opportunities.
Key Lesson: Under the Model Context Protocol, not everything is a "Tool." Differentiating between data (Resources), actions (Tools), and instructions (Prompts) makes your AI architecture secure and efficient.
🧠 Core Concepts: Deep Dive into the 3 Primitives
1. MCP Tools — The "DO" Layer 🛠️
An MCP Tool is an executable capability exposed by an MCP server. It allows the AI model to take action or perform complex, targeted queries. Tools typically require defined inputs (schema) and return an execution result.
- Search Salesforce opportunities based on specific criteria.
- Create a new Salesforce Task or Lead.
- Update the status of an approved record.
- Query a backend inventory database.
- Send an email or Slack notification.
Example: A tool named search_opportunities requires inputs like owner_id and minimum_value. Claude provides these inputs, the tool runs, and Claude analyzes the returned data.
2. MCP Resources — The "KNOW / CONTEXT" Layer 📄
An MCP Resource represents contextual, file-like, or read-only information made available through a specific Resource URI.
- A static customer profile or account history.
- A PDF document or knowledge base article.
- Database schema definitions.
- Product catalogs or pricing sheets.
Example: customer://ABC-10025. Resources provide pure context. Providing a resource does not give the AI permission to change or act on that data.
3. MCP Prompts — The "GUIDE" Layer 📝
An MCP Prompt is a reusable message template that a connected client (or user) can invoke by name. Prompts standardize how common workflows are initiated.
- Code review checklists.
- Incident summary structures.
- Executive briefing formats.
Example: A prompt named customer_review takes arguments like "Customer Name" and "Review Period," ensuring the AI always formats the output exactly how your sales team likes it. Note: Prompts provide instructions, not the actual business data (that comes from Resources or Tools).
🧩 How They Work Together
User: "Prepare me for my meeting with ABC Healthcare." ↓ PROMPT: Injects the "Customer Review" template. ↓ RESOURCE: Attaches the static "ABC Healthcare" company context. ↓ TOOL: Retrieves live, active opportunities from Salesforce. ↓ CLAUDE (LLM): Analyzes the combined data and reasoning. ↓ RESPONSE: Generates the finalized Executive Briefing. ↓ OPTIONAL TOOL: Creates a "Meeting Prep Completed" task in Salesforce.
Rule: Use the right MCP primitive for the right job. Don't use a Tool when a static Resource will do.
Gain: Cleanly separates actions, information, and reusable guidance. Makes debugging and AI reasoning much more accurate.
Price: Requires careful architectural planning. Exposing multiple capabilities means tighter security and governance integration is required.
Limits: A Resource cannot execute actions. A Prompt does not magically fetch business data. Tools must be strictly scoped with least-privilege permissions.
Mirror: Tool = DO | Resource = KNOW | Prompt = GUIDE
Connects to: Claude reasoning, AI Agents, MCP Servers, Agentforce, and Enterprise workflows.
❌ Wrong: "Everything exposed by an MCP Server is called a Tool."
✅ Correct: "MCP servers can expose Tools, Resources, and Prompts. They serve different architectural purposes."
❌ Wrong: "I'll use a Resource to update the Salesforce record."
✅ Correct: "A Resource provides read-only context. To update a record, you must invoke a Tool."
❌ Wrong: "The Prompt contains all the customer's data."
✅ Correct: "The Prompt provides the reusable instructions/template. The actual customer data is injected via Resources or Tools."
💡 Core Q&A
Q: What is the simplest way to remember Tools, Resources, and Prompts?
A: 🎯 Say this first: Tool = DO (Executable capability). Resource = KNOW (Contextual information). Prompt = GUIDE (Reusable message template).
Q: Can one single AI workflow use all three primitives at the same time?
A: Yes. A Prompt can set the rules and guide the workflow, a Resource can provide background knowledge, and a Tool can fetch live database metrics or perform a final action.
🔗 Connecting the Dots (Follow-up Scenarios)
Q: A user types, "Create a follow-up task for ABC Healthcare." Which primitive is most relevant?
A: A Tool. Creating a task is an actionable, executable operation that changes system state.
Q: The AI needs to read the company's official "Code of Conduct" before generating an email. Which primitive is relevant?
A: A Resource. The document is static, read-only contextual information.
Q: The sales VP wants every account review generated by the AI to follow the exact same 3-paragraph structure. Which primitive helps here?
A: A Prompt. This acts as a reusable template to standardize the output format.
Q: How does this relate to Salesforce's Agentforce?
A: While Salesforce Agentforce has its own native Actions and Prompt Builder, MCP provides an open standard. If you are building a custom Anthropic Claude application outside of Salesforce, you use MCP Tools and Resources to safely connect Claude back into Salesforce's APIs.
Q: From a security perspective, why separate Tools from Resources?
A: Risk management. Resources are inherently safer because they are read-only. Tools carry execution risk (like deleting a record or sending an email). By separating them, you can grant an AI broad access to read Resources, while applying strict, human-in-the-loop security gates to Tools.
Tools (DO) + Resources (KNOW) + Prompts (GUIDE)
...are fed into Claude's Reasoning Engine
...which empowers an AI Agent
...to complete an Enterprise Workflow.
In the Salesforce world: Resource = Customer Context, Tool = Update/Create Record, Prompt = Standard Layout.
Official Reference Note: This architecture overview aligns with the latest official Model Context Protocol (MCP) documentation by Anthropic. Because MCP is rapidly evolving, always check the latest SDK protocol versions when implementing live enterprise integrations.
Comments
Questions, corrections and your own experience are all welcome.
Post a Comment