AI/ES/25/ws/1/Building Agents with Model Context Protocol

Properties

created-by: Person/Mahesh Murag
tags: Anthropic/Model Context Protocol, MCP
alias: AIES 25 WS 1 - Building Agents with MCP - Mahesh Murag

Building (Agents) with Model Context Protocol
- About
  - This workshop, from the creators of MCP, talks about the philosophy behind MCP, its impact on the broader ecosystem since launch, and how developers can use it to build context-rich AI apps and agentic experiences.
  - Video and Transcript
    - AI/ES/25/ws/1/Building Agents with Model Context Protocol/YouTube
  - Pics
    - AI/ES/25/ws/1/Building Agents with Model Context Protocol/Pics
  - Summary and Context
    - Latent Space/Blog/25/03/Why MCP Won
- Before Pic 01
  - Pic 01
    My Notes
    - On my left was Person/Mike Christensen, VP Engineering at Fundrise, with quite a bit of experience in Real Estate. Was using IDE/Cline (or was it Windsurf?) to do some Vibe Coding a bit. We talked about MCP and the importance of #Security in Agentic Systems.
    - On my right was Person/Hayden Harrow, Co-Founder at Headless Agents (headlessagents.ai). Mike and I got a demo of his product, which allows deploying any type of agent and allowing inter-agent communication. Right now the product is has developers as its target audience.
- intro
  - what is MCP
    - APIs
      - web applications interact with backend
    - LSP - language server protocol
      - specifies how IDEs to interact with language-specific tools
    - MSP
      - standardize how Ai apps interact with external systems
  - without mcp - fragmented AI development
    - EACH AI app -> custom implementation custom prompt logic -> custom tool calls -> custom data access
    - if different teams in company are doing this
  - MCP - standardized AI development Pic 02
    - AI Notes
      - Any MCP Client:
        • Claude for Desktop
        • Cursor
        • Goose
        • Windsurf
        • Your Client
      - MCP Servers and Their Functions:
      - Database MCP server
        • Query and fetch data
        • Modify records
        • Handle data streams
        → Databases
      - CRM MCP server
        • Create/update contacts
        • Manage opportunities
        • Track interactions
        → CRM systems
      - Version Control MCP server
        • Commit changes
        • Resolve conflicts
        • Manage branches
        → Version Control Software
    - My Notes with mcp - you build it once and get adoption everywhere
      - for AI app devs
        connect your app to any mcp server with
      - for tool or API devs
        build mcp server once, see adoption everywhere
      - for end users
        more powerful and context-rich ai applications
      - for enterprises
        clear separation of concerns between teams
        #Insight
        you can have one team that owns the relationship to the vector database
  - fast adoption Pic 03
    - Pic 03
    - AI Notes fast adoption
      - Companies & Projects:
        Cloudflare
        Replit
        Shopify
        IBM
        GitHub
        Block
        Grafana
        Sourcegraph
        Cursor
        Docker
        Codeium
    - My Notes
      - 1100 community built servers
      - Person/Jack Dorsey mentioned it yesterday
      - we think this is going to be industry-standard
  - MCP deep dive Pic 04
    - Pic 04
    - My Notes
      - MCP Client
        invokes tools
        queries for
      - server can interpolate
      - resources
        manifest as attachments
        attached to the chat
    - AI Notes
      - MCP Client
        • Invokes Tools
        • Queries for Resources
        • Interpolates Prompts
      - MCP Server
        • Exposes Tools
        • Exposes Resources
        • Exposes Prompts
      - Components:
      - Tools (Model-controlled)
        • Functions invoked by the model:
        • Retrieve / search
        • Send a message
        • Update DB records
      - Resources (Application-controlled)
        • Data exposed to the application:
        • Files
        • Database Records
        • API Responses
      - Prompts (User-controlled)
        Pre-defined templates for AI interactions:
        Document Q&A
        Transcript Summary
        Output as JSON
    - My Notes
      - #Example IDE/Zed AI/Coding Slash Command
      - Q - why aren't resources modeled as tools
        A - it's not about making model better but defining ways for application to interact with the server
        we need separation between model-controlled and not model-controlled
        contract for the client builder and server builder for what should be invoked where
      - langgraph just made adapters to connect to MCP
        if it has this connection to mcp servers you can connect it to your agent
        we don't think mcp is going to replace agent frameworks
      - a few questions are about "why isn't it a tool"
      - a few questions about "why do we need agent frameworks if we need this"
      - #Q - why do resources and prompts exist, because you can do it with tools
        a - there are more capabilities with tools and resources
        in reality, resources and prompts can be dynamic and interpolated with context from the user or the aplication and the server can return a customized resource or prompt for the task at hand
        client can subscribe to resource, and server can notify the client and tell the client that you need to update the state of your system
        you can do a lot with tools, but MCP isn't just about giving the model more context, it's about getting richer ways to give the application
  - MCP in Action Demo Pic 06
    - My Notes
      Claude/Desktop
      give url of github repository
      the model, claude, does the invoke issues tool, and pulls those into context, summarizing them
      it's personalizing to you based on your history with claude and what it knows you want
      who you are,
      what you are working on
      how you like to interact with it
      "can you triage the top 3 priority issues and add it to my asana project"
      doesn't give name of asana
      has installed asana server
      30 tools
      finds the projects
      starts adding these issues as tasks
      callouts
      I didn't build the github or asana server
      each are a couple hundred lines of code
      it's not a ton of additional work to build though
      they are all playing together with claude for desktop
      (video)
    - AI Notes
      MCP in Action
      Left: Claude AI Ticket Prioritization
      Lists various technical issues:
      Prompt Caching Support for Vertex/Bedrock (#774, #653)
      Multiple requests for prompt caching support
      Parity gap between direct API and cloud provider integrations
      Could improve performance and reduce costs
      PDF File Support Issues (#725)
      Problems with PDF document handling on Bedrock
      Impacts users leveraging Claude's document understanding
      Testing Framework Updates (#647)
      Test suite failures with pytest 8
      Maintaining test reliability is important for development
      
      Prompt in UI:
      "Can you triage the top 3 highest priority issues and add them to my Asana project for SDK Fixes?"
      Right: GitHub Issues Page (Anthropic SDK)
      Open issues list:
      Some API error codes don’t appear to be supported in SDK (#853)
      OpenAI Compatibility using Claude API Key (#851)
      Feature request: tool_choice="none" to prevent unintended tool calls (#849)
      Caching seems to be enabled by default (#837)
      
      Speaker: Vibhu Sapra
      Presented by Anthropic
  - JSON Task Plan in VS Code Pic 07-
    - Pic 07
    - My Note
      Why is this JSON format in a markdown file?
    - AI Notes
      File: plan.md
      Description: Verify and validate collected information
      Tasks:
      Cross-reference technical claims about quantum computing capabilities and encryption threats
      Agent: fact_checker
      Verify timeline predictions and industry statistics for consistency across sources
      Agent: fact_checker
      Validate technical details of proposed quantum-resistant solutions
      Agent: fact_checker
  - MCP-Agent: Orchestrator Setup in Python Pic 08
    - Pic 08
    - AI Notes
      - Sidebar - Project Structure (VS Code project explorer on left)
        mcp-agent ├── output/ ├── sources/ │ ├── kyber_research.txt │ ├── nist_pqc_standard.txt │ ├── nist_quantum_cr.txt │ ├── palo_alto_quantum.txt │ ├── quantum_security.txt ├── main.py ├── mcp_agent.config.yaml ├── mcp_agent.secrets.yaml ├── plan.md ├── README.md ├── requirements.txt ├── task.md
      - Code: main.py (on right)
        python
        async def example_usage(): server_names = ["filesystem", "fetch"] # Read the research task task_path = Path("task.md") with open(task_path, "r") as f: task = f.read() # Create orchestrator with all agents orchestrator = Orchestrator( llm_factory=AnthropicAugmentedLLM, available_agents=[ search_agent, fact_checker, report_writer, ], plan_type="full", plan_output_path=Path("output/execution_plan.md"), max_iterations=5 ) # Execute the research task result = await orchestrator.generate_str( message=task, request_id="20241022" )
      - Key Features
        Uses AnthropicAugmentedLLM as the LLM factory
        Agents involved:
        search_agent
        fact_checker
        report_writer
        Reads tasks from task.md
        Outputs execution plan to "output/execution_plan.md"
        Runs up to 5 iterations
    - My Notes
      - Here's Windsurf
        they have their own ui inside of their agent
        other applications don't call them tools (one calls them extensions)
        -> which IDE is that? booz?
      - his claim: #MCP will be "the foundational protocol for agents"
        the models are becoming better and they use the data that you bring to them in better ways.
  - Anthropic/Blog/24/12/Building Effective Agents with #MCP 1 - Pic 09
    - Pic 09
    - My Notes
      - now, with MCP
      - the idea of an augmented LLM
      - takes inputs and outputs
      - LLM
        query and results with retrieval systems
        call and response with tools
        read and write with memory
      - one of the key points of MCP is that the system can discover new capabilities after the agent system is built
      - agent is an "augmented LLM" running in a loop
        working towards
      - you can let users of the agent system connect to their data and you as the developer can focus on the core loop
        the agent can be focused with the interaction with the llm as it's core
    - AI Notes
      - Sampling →
        Allows a server to request completions from a client, giving the user application full control over security, privacy, and cost.
      - Client-Server Interaction
        Client ←→ Server (Request for inference)
      - Sampling Parameters:
        Model Preferences & Hints
        System Prompt
        Temperature & Max Tokens
  - what does this look like in practice - MCP/mcp-agent
    - mcp-agent is a framework built by lastmile.ai
      - MCP/mcp-agent
        perhaps - lastmile-ai/mcp-agent: Build effective agents using Model Context Protocol and simple workflow patterns
    - case study -> research task -> quantum computing impact on cybersecurity
    - sub-agents
      - research agent
        look up urls
      - fact checker agent
        verify info coming in
        brave, fetch, filesystem tools
        each are mcp servers that it has access to
      - research writer
    - we get a plan.md and a task.md - Pic 07
      - Pic 07
      - Screenshot 2: GitHub Issues Page (Anthropic SDK)
        Open issues list:
        Some API error codes don’t appear to be supported in SDK (#853)
        OpenAI Compatibility using Claude API Key (#851)
        Feature request: tool_choice="none" to prevent unintended tool calls (#849)
        Caching seems to be enabled by default (#837)
    - My Notes
      - to close out this part of demo
        it starts doing some research
        invoking the fact checking agent
      - lastmile-ai/mcp-agent: Build effective agents using Model Context Protocol and simple workflow patterns
        someone asks about this and what's adapted
      - question - how do resources and prompts fit in in this case
        resources and prompts come in when the user is in the loop
        what if you had a nice loop in an extended chat interface
        the chat interface application could take this plan and surface it to you
        aplpication could have nice UI on top that shows the plan beautifully in the UI
        prompts example
        slash command - slash summarize (predefined prompt in server to give the user a summary)
      - questions - how does #MCP fit into evaluations
        largely this is the same as it is right now
        this could be a standard layer inside of evals themselves
        there could be an MCP server that has the same 5 tools
      - question - I'm missing the infrastructure side of this
        are these servers exposed globally, or only locally?
        comment: line between client and server not clear to one person
        how do you do authentication, retry logic, etc
        q - what is the separation between the logic, retry logic, authentication
        for retry logic, personal opinion is that it should happen on the server side
      - Q/My how do we do versioning well with MCP servers
      - #Q is there a best practice or a limit to the number of servers you can expose to an LLM?
        today, they are good up to 50 or 100
        Claude is good up to 100 in my opinion
        but how do you avoid overloading context
        "a tool to search tools" is exciting
        RAG over tools
        the entire library of tools that are
        hierarchical systems of tools
        #Observation this reminds me of that LangChain/Blog/25/02/Benchmarking Single Agent Performance
      - #Q - they have an entire page on how to make mcp server with claude or llms
        he wrote all of those with claude
        it's really easy to approach it
        start with tool type
      - #Q - if these are simple, can lLms gerentate
        if you've heard of cline, they have an mcp auto generator tool
        if you want to start talking to gitlab
        that works for simple servers
        but if you want logging or data traan
      - #Q are you talking to the owners of the services and the data
        a lot of th servers are public already
      - Q/My - how do you handle versioning of tool descriptions and tools available
        typescript servers have semvar version
        no best practice yet
  - Building Effective Agents with MCP 1 - cont'd - Pic 09
    - My Notes
      - in typical applications, client invokes a server, but a class of server exists where the server will
      - server asks the user for more information so you want the server to get intelligence
      - federate these requests
      - client doesn't have to listen to it
        has control over cost or privacy
        limit certain # of requests
      - one of the design principles
    - AI Notes
      - Building Effective Agents with MCP
      - Sampling →
        
        Allows a server to request completions from a client, giving the user application full control over security, privacy, and cost.
      - Client-Server Interaction
        arduino CopyEdit Client ←→ Server (Request for inference)
      - Sampling Parameters:
        Model Preferences & Hints
        System Prompt
        Temperature & Max Tokens
  - Building effective Agents with MCP - 2 - Pic 10
    - AI Notes
      - Building Effective Agents with MCP
      - Composability →
        Key Insight An MCP/Server can be a server and vice-versa.
      - Client-Server Chain
        User & LLM ←→ Client/Server ←→ Client/Server ←→ Client/Server
      - The architecture allows for recursive composition where MCP clients can also act as servers.
      - Enables modular AI workflows by chaining multiple MCP instances.
    - My Notes
      - microservices and chaining
      - Q - how do you deal with compounding errors if the system itself is complex and multi-layered
        same dynamic as with multi-layer agents
        it's up to each successive layer to deal with controlling data as it's structured
        the third node up there should collect data and fan in data from all the other ones and make sure it's
      - #Q - How do you do observability and how does the first client/server know about the results of the 3rd hop client/server?
        it's you don't exactly know what's behind the hood
        it's kind of a black box
      - #Q - how do you make #MCP able to be #Debuggable
        it's up to the devs right now
      - #Q this is similar to Microservices
        best practices from there should apply
      - #Q what if client wants control over the tool call
        one suggestion is doing that via a prompt
        they are thinking about something called MCP/Tool/Annotations to limit the number of tools to 5
      - #Q - what #Debugging patterns have you seen
        they have something called MCP/Inspector
        he's seen servers
      - #Q - how are you thinking about #Security #Good Question
        we're going to bake in Auth
what's coming next
- Building effective agents with #MCP - 3 - Pic 11
  - - Notes from Hayden - the tool specs are sent back, LLM is only called on client
  - AI Notes
    - Building Effective Agents with MCP
    - Sampling + Composability
      
      Chain agents together while ensuring that the client application controls inference.
    - Architecture
      arduino CopyEdit Application + LLM → Client/Server (Orchestrator Agent) ├── Client/Server (Analysis Agent) ├── Client/Server (Coding Agent) ├── Client/Server (Additional Agent)
    - The Orchestrator Agent acts as the central hub, distributing tasks to specialized agents.
    - Each agent functions as both a client and server, supporting modular and scalable AI workflows.
    - Enables controlled inference with hierarchical delegation.
  - My Notes
    - orchestrator agent communicates with and fans out to analysis, coding, and research agents
    - we need guarantees about connectivity
  - MCP/Inspector with #Auth Pic 12
    - AI Notes
      - MCP Inspector with Auth!
        MCP Inspector v0.4.0
        Transport Type: SSE
        URL: (Redacted)
        Connection Status: ✅ Connected
        Resources Panel
        List of resources:
        general
        random
        justin-testing
        test
        slack-connect-testing
        justin-testing-2
        Resource Templates Panel (Empty)
        Displayed JSON Response (Slack Message)
        json
        { "contents": [ { "uri": "slack://conversations/C04LN9C1SL/messages", "mimeType": "application/json", "text": "{\"channelId\":\"C04LN9C1SL\",\"messages\":[{\"text\":\"Unsubscribed from <https://github.com/modelcontextproto>\"}]}" } ] }
  - MCP Inspector with Slack OAuth Authentication - Example OAuth flow
    - My Notes
      - SSE is the "best way" to do MCP/Servers that are remote
      - #Example
        #Slack auth
        the highlight here
        num 1 thing -> remotely hosted servers
        discoverable
        you don't have to mess with stdio
        you can
        this is going to be a big explosion in the number of servers
        you as a user don't need to know how to build it or host it
        #Q what about elevating permissions
        #Q the server then holds the token -> isn't that a bad thing?
        if you think about the server being the one that's closest to slack
        let's say slack builds that and then slack will want to control that
        #Q does MCP replace REST in the long term?
        LLMs need data in a different way than the server
        RESTful is for stateless interactions
        #Q - when we wrap our tools with MCP, how do we deal with evals, how do we think about when tools change and regressions happen
        we are going to talk about the registry
        this doesn't change too much about the eval system around tools
        fixtures and diffs in the evals
        #Q how far away in the spec is it
        days away
    - AI Notes
      - MCP Inspector with Slack OAuth Authentication
      - Slack OAuth Prompt
      - Slack MCP Server is requesting permission to access the Development Environment Slack workspace
      - Requested Permissions:
        Content and info about you
        Content and info about channels & conversations
      - Options:
        [Cancel]
        [Allow]
- MCP/Registry/API #API - 1 - Pic 14
  - My Notes
    - this lets us have something like npm, pypi, Package/Managers for mcp
    - how do you determine whether it's sse or stdio
    - who built it, is it verified, did Shopify bless this server?
    - "this is coming, it's going to be great" - #Quote from Person/Mahesh Murag
    - #Q can you give more examples of the server calling the client?
      - they don't yet have the ability to do server side "sampling" MCP/Sampling
        Q/My what is sampling again here?
    - #Q does the model have to be in the loop with the LLM
      - someone asks this ... didn't hear response
- An Official MCP Registry API - Pic 15
  - Pic 15
  - My Notes
    - agents can do their own discovery - agent will find out what it's capable of
  - AI Notes
    - An MCP server registry helps make agents self-evolving by letting them discover and choose their own tools.
    - Workflow Example: Debugging with Grafana
      pgsql CopyEdit User → "Fix the bug based on my Grafana logs" ↓ Agent → Queries Grafana logs & suggests a fix ↓ Agent → Searches for official Grafana server ↓ Agent → Installs & invokes Grafana server ↓ Registry API
    - Key Idea: Agents can autonomously locate and use the necessary tools by querying the MCP Registry API.
    - Enables adaptive AI workflows where agents evolve based on available tools.
- Server discovery - Pic 16
  - My Notes
    - #Example - #Shopify
    - registry is focused on discovery and ability to find tools from scratch
    - if you know you want to talk to shopify, it can check this MCP/.well-known - really nice complement to AI/Computer Use.
    - what if you could have that PLUS MCP json (?)
      - didn't quite get this point
  - AI Notes
    - Server Discovery with MCP
    - Workflow Example: Managing a Shopify Store
      - User → "Help me manage my store on Shopify" ↓ Agent → Checks if Shopify has a `.well-known/mcp.json` ↓ Agent → Updates store on Shopify via Remote MCP server
    - Shopify MCP Server Metadata
      - json
        { "version": "1.0", "servers": [ { "id": "main-server", "name": "Shopify Store Access", "endpoint": "https://shopify.com/api/mcp", "capabilities": ["resources", "tools"], "authType": "oauth2" }, { "id": "admin-server", "name": "Shopify Admin API", "endpoint": "https://admin.shopify.com/api/mcp", "capabilities": ["resources", "tools"], "authType": "oauth2" } ], "organization": { "name": "Shopify" }, "requirements": { "minProtocolVersion": "0.5.0" } }
    - Key Features
      - Uses .well-known/mcp.json for automatic service discovery.
      - Supports OAuth2 authentication.
      - Defines multiple MCP endpoints (Store Access, Admin API).
      - Ensures compatibility using minProtocolVersion.
- And a lot more - Pic 17
  - My Notes
    - in order of immediacy
    - they are using something like SSE
      - streaming -> stream data from server to client
    - elicitation -> "proctive" event behavior
      - better patterns for that.
  - AI Notes
    - And a lot more...
    - Long-lived vs short-lived connections
      Smooth upgrade from stateless to stateful capabilities
    - Streaming
      Stream data from server to client
    - Namespacing
      Prevent collisions between entities of multiple servers; create logical groups of servers or tools
    - Elicitation
      Servers proactively request context from users
Q/My
- how exactly do permission boundaries work with #MCP ? he mentions OAuth which has some concept of permission boundaries ...

AI/ES/25/ws/1/Building Agents with Model Context Protocol

About

Video and Transcript

Pics

Summary and Context

Before Pic 01

intro

MCP - standardized AI development Pic 02

My Notes with mcp - you build it once and get adoption everywhere

fast adoption Pic 03

AI Notes fast adoption

Companies & Projects:

MCP deep dive Pic 04

resources

MCP Client

MCP Server

Tools (Model-controlled)

Resources (Application-controlled)

Prompts (User-controlled)

#Example IDE/Zed AI/Coding Slash Command

#Q - why do resources and prompts exist, because you can do it with tools

MCP in Action Demo Pic 06

MCP in Action

Left: Claude AI Ticket Prioritization

Right: GitHub Issues Page (Anthropic SDK)

JSON Task Plan in VS Code Pic 07-

File: plan.md

MCP-Agent: Orchestrator Setup in Python Pic 08

Sidebar - Project Structure (VS Code project explorer on left)

Code: main.py (on right)

Key Features

Here's Windsurf

his claim: #MCP will be "the foundational protocol for agents"

Anthropic/Blog/24/12/Building Effective Agents with #MCP 1 - Pic 09

Sampling →

Client-Server Interaction

Sampling Parameters:

what does this look like in practice - MCP/mcp-agent

Screenshot 2: GitHub Issues Page (Anthropic SDK)

Building Effective Agents with MCP 1 - cont'd - Pic 09

Building Effective Agents with MCP

Sampling →

Client-Server Interaction

Sampling Parameters:

Building effective Agents with MCP - 2 - Pic 10

Building Effective Agents with MCP

Composability →

Client-Server Chain

Q - how do you deal with compounding errors if the system itself is complex and multi-layered

#Q - How do you do observability and how does the first client/server know about the results of the 3rd hop client/server?

#Q - how do you make #MCP able to be #Debuggable

#Q this is similar to Microservices

#Q what if client wants control over the tool call

#Q - what #Debugging patterns have you seen

#Q - how are you thinking about #Security #Good Question

what's coming next

Building effective agents with #MCP - 3 - Pic 11

Building Effective Agents with MCP

Sampling + Composability

Architecture

MCP/Inspector with #Auth Pic 12

MCP Inspector with Auth!

MCP Inspector v0.4.0

Resources Panel

Displayed JSON Response (Slack Message)

MCP Inspector with Slack OAuth Authentication - Example OAuth flow

#Q what about elevating permissions

#Q the server then holds the token -> isn't that a bad thing?

#Q does MCP replace REST in the long term?

#Q - when we wrap our tools with MCP, how do we deal with evals, how do we think about when tools change and regressions happen

#Q how far away in the spec is it

MCP Inspector with Slack OAuth Authentication

Slack OAuth Prompt

MCP/Registry/API #API - 1 - Pic 14

#Q can you give more examples of the server calling the client?

#Q does the model have to be in the loop with the LLM

An Official MCP Registry API - Pic 15

Workflow Example: Debugging with Grafana

Server discovery - Pic 16

Server Discovery with MCP

File: `plan.md`

Code: `main.py` (on right)