In 2026, MCP has become the de facto enterprise AI connectivity layer, with implementations across Claude Cowork, Claude Code, and a growing list of third-party AI tools. If your company is deploying Claude Cowork, adopting enterprise AI, or evaluating AI connectivity standards, you need to understand MCP.
What is MCP, in one paragraph?
MCP (Model Context Protocol) is an open, JSON-RPC-based protocol that standardizes how AI models access external tools, data sources, and services. Instead of every AI vendor building custom integrations for every system, MCP provides a single interface: the system exposes an “MCP server,” and any MCP-compatible AI can connect to it. Think of MCP as USB-C for AI — one protocol, many implementations, infinite compatibility.
Why MCP exists — the problem it solves
Before MCP, connecting a large language model to an enterprise system required custom work for every pairing. If you wanted Claude to read your Salesforce data, someone wrote a Claude-specific Salesforce integration. If you wanted ChatGPT to read it too, someone wrote a second one. Every vendor, every system, every combination was bespoke.
MCP solves this by specifying:
- A standard server interface — any system can expose tools, resources, and prompts through the same JSON-RPC spec
- A standard client — any AI application can consume any MCP server
- A standard auth model — OAuth 2.1 flows, bearer tokens, capability negotiation
- A standard discovery mechanism — clients can query a server to learn what tools and resources are available
How MCP works — the architecture
The three concepts
- Tools — Actions the AI can invoke:
create_ticket,search_customers,update_deal_stage. - Resources — Data the AI can read. A SharePoint document, a database row, a file. Resources have URIs and content types.
- Prompts — Reusable prompt templates the server exposes.
The communication pattern
MCP uses JSON-RPC 2.0 over either standard I/O (for local servers) or HTTP with Server-Sent Events (for remote servers). An AI client initializes a connection, negotiates capabilities, discovers available tools and resources, then calls them as needed during a task.
[ Claude Cowork (client) ]
↕ JSON-RPC
[ MCP Server for Salesforce ]
↕ Salesforce REST API / OAuth
[ Your Salesforce Org ]The MCP server is a thin adapter layer. It translates your system’s native API into the MCP spec. Deploy it on a managed cloud host, in your VPC, or on the user’s machine.
MCP vs. REST APIs — why not just use the API?
Technically, a sufficiently advanced AI could call any REST API directly. In practice, MCP solves four problems REST doesn’t:
- Discovery. MCP servers can be queried at runtime: “what tools are available? What are their schemas?”
- Uniform auth and session. MCP specifies OAuth 2.1 flows and session management once, for all systems.
- Streaming and long-running operations. MCP is designed for AI workflows: partial results, progress events, cancellation.
- Prompt and resource primitives. REST has no equivalent to a “resource” the model can browse.
What an MCP connector actually does (with examples)
Example 1 — Salesforce MCP connector
- Tools:
search_accounts,get_account_detail,get_opportunity,create_note,update_opportunity_stage - Resources:
salesforce://accounts/{id},salesforce://opportunities/open - Auth: OAuth 2.1 with Salesforce Connected App. Build time: 4–6 days.
Example 2 — SharePoint MCP connector
- Tools:
search_documents,get_document,list_sites,get_permissions - Auth: Azure AD / Entra via Microsoft Graph OAuth. Build time: 5–8 days.
Example 3 — Internal API MCP connector
Tools bespoke to the system. Build time: highly variable. Clean modern APIs: 3–5 days. Legacy SOAP / on-prem: 2–4 weeks.
Common enterprise MCP use cases
- CRM connectivity — Salesforce, HubSpot, Dynamics
- Document management — SharePoint, Google Drive, Dropbox Business
- Ticketing and PM — Jira, Linear, ClickUp, Asana
- Support platforms — Zendesk, Intercom, Freshdesk
- HRIS — Workday, BambooHR
- Internal APIs — your microservices, custom databases, legacy systems
- Analytics — Mixpanel, Amplitude, Looker
- Communication — Slack, MS Teams
What MCP connectors cost to build
| Connector type | Complexity | Typical cost range |
|---|---|---|
| Public SaaS, good API | Low | Low four figures |
| Enterprise SaaS | Medium | Mid four figures |
| Microsoft Graph / SharePoint | Medium-High | Mid-to-upper four figures |
| Internal API, custom auth | High | Upper four to low five figures |
| Legacy / on-prem / SOAP | Very High | Low-to-mid five figures |
Build vs. buy — how to decide
Build in-house when:
- You have engineers with TypeScript/Python experience and time
- The system is proprietary/internal
- The logic is deeply specific to your workflows
Buy or commission when:
- Speed matters
- The system is standard SaaS
- Your engineers’ time is better spent on product
MCP and Claude Cowork — the integration story
MCP is the plumbing that makes Claude Cowork actually useful at work. Without MCP connectors, Claude Cowork is a smart document reader. With them, Claude Cowork becomes an operator across your stack.
Anthropic has first-class MCP support baked into Claude Cowork:
- Admin console for managing connected MCP servers
- Per-user permission scoping
- Audit logs for every MCP tool call
- Rate limiting and cost attribution
See how we build MCP connectors →
Frequently asked questions about MCP
Need an MCP connector built?
Tinkso builds custom MCP connectors for mid-market Claude Cowork deployments. Fixed-fee, production-grade, typically shipped in 3–10 days per connector.