Building Agents with Model Context Protocol - Full Workshop with Mahesh Murag of Anthropic
by AI Engineer (Mahesh Murag, Anthropic)
The workshop that taught the ecosystem what MCP is, from Anthropic's Applied AI team.
Overview
Recorded for the AI Engineer channel on 1 March 2025 and delivered by Mahesh Murag of Anthropic's Applied AI team, this 1 hour 44 minute workshop is the reference video introduction to the Model Context Protocol. It opens on the integration problem MCP exists to solve — the N-times-M explosion of bespoke connectors between AI applications and the tools they need — and argues for a common layer between application developers and tool providers. It then works through the protocol's three primitives and, importantly, who controls each: tools, where the model decides when to invoke; resources, which the application governs; and prompts, which the user selects. A demonstration section shows MCP running in Claude for Desktop and against a GitHub integration. The talk then turns to how MCP relates to existing agent frameworks, arguing it complements rather than replaces them by giving an augmented LLM a way to discover capabilities after initialisation. Later chapters cover the capabilities that matter most for agents — sampling, where a server asks the client for a model completion, and composability, where a server is itself a client of other servers — before closing on the roadmap as it stood in early 2025: remote servers, authorisation, an official registry, self-evolving agents that discover tools at runtime, and proactive server behaviour. A long audience Q&A on controllability, model choice and server proactiveness ends the session.
At a Glance
- Topic
- MCP
- Level
- All Levels
- Format
- Video
- Cost
- Free
- Duration
- 1 hour 44 minutes
- Provider
- AI Engineer (Mahesh Murag, Anthropic)
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓Why MCP exists: the N-times-M integration problem and the common-layer answer
- ✓The three primitives and who controls each — tools, resources and prompts
- ✓How MCP complements rather than replaces existing agent frameworks
- ✓What sampling is and why a server asking the client for a completion matters
- ✓How composability lets a server act as a client of other servers
- ✓Watching MCP work end to end in Claude for Desktop and against GitHub
- ✓How dynamic capability discovery changes agent design after initialisation
- ✓The protocol roadmap Anthropic set out: remote servers, auth and a registry
Highlights
- •Delivered by Anthropic's own Applied AI team, so the design intent is first-hand rather than inferred
- •The most-referenced video introduction to MCP, with independent chapter-by-chapter notes published by practitioners
- •Explains the control model — model-controlled, application-controlled, user-controlled — which most tutorials skip and which is the part teams get wrong
- •Goes past the basics into sampling and composability, the two features that actually make MCP interesting for agents
- •Dated deliberately: the March 2025 roadmap section is now history — remote servers, OAuth and the official registry all shipped — so treat the last half hour as context, not a preview
Who It’s For
Best For
- ✓Engineers who want an authoritative conceptual grounding in MCP before coding
- ✓Developers deciding whether MCP or a framework abstraction fits their agent
- ✓Team leads who need to explain the protocol's value to other engineers
- ✓Anyone who finds a two-hour talk easier than reading a specification
Prerequisites
- •Familiarity with LLMs and basic agent concepts such as tool calling
- •Some programming experience; no MCP experience required
- •Read the current MCP specification alongside it, since the roadmap section is from early 2025
FAQ
What is Building Agents with Model Context Protocol - Full Workshop with Mahesh Murag of Anthropic?
A 1 hour 44 minute AI Engineer workshop from March 2025, delivered by Mahesh Murag of Anthropic's Applied AI team, covering MCP from the integration problem it solves through its three primitives to sampling, composability and the protocol roadmap. For engineers who want the authoritative conceptual grounding before writing a server.
Is Building Agents with Model Context Protocol - Full Workshop with Mahesh Murag of Anthropic free?
Building Agents with Model Context Protocol - Full Workshop with Mahesh Murag of Anthropic is free to access.
What level is Building Agents with Model Context Protocol - Full Workshop with Mahesh Murag of Anthropic for?
Building Agents with Model Context Protocol - Full Workshop with Mahesh Murag of Anthropic is aimed at a all levels audience. Recommended background: Familiarity with LLMs and basic agent concepts such as tool calling, Some programming experience; no MCP experience required, Read the current MCP specification alongside it, since the roadmap section is from early 2025.
How long does Building Agents with Model Context Protocol - Full Workshop with Mahesh Murag of Anthropic take?
Expect roughly 1 hour 44 minutes. Most learners work through it at their own pace.
What will I learn from Building Agents with Model Context Protocol - Full Workshop with Mahesh Murag of Anthropic?
You'll learn: Why MCP exists: the N-times-M integration problem and the common-layer answer; The three primitives and who controls each — tools, resources and prompts; How MCP complements rather than replaces existing agent frameworks; What sampling is and why a server asking the client for a completion matters; How composability lets a server act as a client of other servers; Watching MCP work end to end in Claude for Desktop and against GitHub; How dynamic capability discovery changes agent design after initialisation; The protocol roadmap Anthropic set out: remote servers, auth and a registry.
Topics
Sources
This page was written from 3 sources, 2 on domains other than youtube.com.