← All sessionsHomeSearch
C7 EST | 14 Day AI Sprint·Day 5 | Gen AI Beyond Prompting: Introduction to MCPs·4:59:00

Day 5: MCP as Concierge — Easy/Medium/Hard Modes, Stacking Servers, Smithery Vetting, and RAG vs MCP

Dileep Day mentor - Head of Generative AI Initiatives at Outskill; PhD researcher (IIT Madras, ~8 years, international publications); teaches MCP as three difficulty modes with live Claude/ChatGPT demos · Harshith Office-hours co-mentor - 'MCP is Harsha's bread and butter'; answers n8n, RAG and MCP questions · Cameron Office-hours co-mentor - agency owner; vibe-coding, lead-gen and business questions · Uthappa Host - recap of days 1-4, breaks, weekend schedule
▶LMS lesson

The short version

  1. MCP (Model Context Protocol) defined by analogy three times: API = you keep checking whether the food is ready, webhook = the kitchen calls you, MCP = a concierge handles the whole exchange for you. The 'universal travel adapter' between any app and an LLM - and by definition an LLM must be in the loop, so two apps talking in n8n is not MCP.
  2. Three modes. EASY: connectors built into ChatGPT/Claude web - no install, any device, but paid plans only (~$20/mo) and MCP context burns your usage limits faster ('deliberate monetization design'). MEDIUM: Claude Desktop + Node.js unlocks ~30 Desktop Extensions on the free plan but is device-bound; Context7 is 'always' his first recommendation. HARD: Smithery.ai's 3,700+ community servers, vetted by quality >60/100, >1,500 monthly calls, uptime >60%, recent deploy.
  3. The real power is STACKING: one prompt that sequences servers - File System reads ds_salaries.csv -> Claude analyses -> Context7 fetches the charting docs -> a results website. 'Explicitly mention the MCP server flow in the prompt' or the model guesses wrong once you have 10+ servers connected.
  4. RAG vs MCP, settled in the office hour: 'RAG is retrieval augmented generation. It's the memory'; MCP 'is executing functions and doing things on your behalf.' Good RAG still needs unstructured-to-structured normalization plus metadata tags; hybrid + semantic retrieval with 'Venn diagram' overlap for the best chunks.
  5. Closing non-MCP segment: Shopify's CEO mandate saw NON-technical teams adopt fastest (a sales team's page-speed comparison tool doubled close rates); Dan Shipper's 7-figure company runs on '100% AI written code' and hired an 'AI Ops Expert' - the program's 'AI generalist'. Three adopter types: resistors, adopters (+10-20%), transformers.

The concepts

01

API, webhook, MCP: check, get called, or send the concierge

You could walk to the kitchen every five minutes. You could ask the chef to shout when it's ready. Or you could send someone whose whole job is handling that for you.

Two paired analogies, each told twice: a food court and a spouse in the kitchen. Checking repeatedly is an API call; being notified is a webhook; a concierge (or your kid) who carries the request, waits, negotiates and returns the result is MCP - 'Model Context Protocol acts like a concierge between you and the LLM.' Formally: Model = the LLM, Context = application data it did not have, Protocol = the connection rules. Positioned as 'the universal travel adapter' - one standard instead of a custom integration per app - proposed by Anthropic and open-sourced.

Two boundary rules: an LLM must be in the equation ('the model is important'), so wiring two apps together in n8n is integration, not MCP; and the 'magic' is giving a model context it was never designed to have - connect Canva and Claude can produce a presentation. The office-hour version: the LLM is Thanos' gauntlet, each MCP server an infinity stone.

Why it matters

It extends Day 1's waiter analogy into the protocol Paul's Catalyst sessions assume as background (Basecamp 5, Session 11 connectors).

MCP by difficulty (Day 5, completed Day 12): install, plan, reach, portability MODE YOU INSTALL PLAN SERVERS WHERE IT WORKS EASY nothing (web connectors) paid ~$20/mo ~30 any signed-in device MEDIUM Claude Desktop + Node.js free OK ~30 more (60+) this machine only HARD Smithery URL or npx free OK 3,700+ this machine; vet 60/1,500/60% ULTRA edit config JSON (LLM merges) free OK 17,433+ or YOUR OWN this machine; stars/forks STACK them: one prompt, servers in a stated order - File System -> analysis -> Context7 -> website Name the server in the prompt once you have more than ~10 connected
Dileep's three MCP modes on Day 5: what you install, what plan you need, how many servers you unlock, and where each is portable.
This idea elsewhereparallelCatalyst Basecamp 5 gives the host/client/server view of the same protocolextendsSame analogy family, one level up
Go deeper

In one line: MCP = open protocol (Anthropic) letting an LLM discover and call external tools/data through servers; a concierge model vs API pull and webhook push; requires an LLM in the loop.

API = you check; webhook = they notify; MCP = concierge does the exchange (l3186027 0:17-0:19)

Model + Context + Protocol; 'universal travel adapter' (l3186027 0:21-0:28)

An LLM must be part of the loop or it is not MCP (l3186027 0:28)

Proposed by Anthropic, open-sourced, adopted by competitors incl. OpenAI (l3186027 0:33; l3302271 0:33)

Gauntlet/infinity-stone analogy: each server adds a power (l3302271 0:21)

▶ Watch this taught:

02

Easy mode: native connectors, paid plans, and the token tax

No install, works from your phone - and it is exactly why the $20 plan exists.

Easy mode is the web-hosted connector menu inside ChatGPT (~30 apps at chatgpt.com/apps: Spotify, Canva, Booking.com, Tripadvisor - live Spotify playlist demo) and Claude (Settings -> + -> Manage Connectors -> Browse -> Web tab: Trivago hotel search for a Muscat trip, Google Drive search). Approve tool calls with 'Allow once' while learning; the model turns the raw API-style payload into a readable table. 'Easy mode is available only on paid plans' - Claude Pro / ChatGPT Plus.

Two cautions. Connector context is not under your control and exhausts usage limits faster - framed as intentional monetization. And the model picks which server to call from each server's stored short description, so once you exceed ~10 connected servers, name the server in the prompt. Side benefit: grounding through a connector 'is a great way to reduce the hallucinations of your tools.'

Why it matters

Easy mode is where a non-developer should start; the token-tax warning explains surprise 'usage limit reached' messages.

This idea elsewhereparallelSame connector walkthrough in Catalyst Basecamp 5
Go deeper

In one line: Easy mode = vendor-hosted connectors in ChatGPT/Claude web; zero install, any device, paid plan required; name the server once you have many.

No installation; web-hosted; any signed-in device (l3186027 0:35)

Paid plans only (~$20/mo) (l3186027 0:36, 1:00)

~30 ChatGPT connectors; Claude path: + -> Connectors -> Browse -> Web tab (l3186027 0:39-0:49)

Connector context burns limits faster - by design (l3186027 1:00-1:03)

Server chosen from stored descriptions; name it explicitly past ~10 servers (l3186027 1:04-1:07)

Easy + medium together ~60+ servers (l3186027 1:27)

▶ Watch this taught:

03

Medium mode: Claude Desktop + Node.js, Desktop Extensions, and why Context7 comes first

Two installers and you are off the paid plan - but the setup lives on this laptop and nowhere else.

Install Claude Desktop and Node.js (prebuilt .msi/.pkg, not source). Node is explained as 'a translator... which tells Claude where MCP servers are and how to activate.' Then + -> Connectors -> Manage -> Browse -> the Desktop Extensions tab (not Web) -> Install. Works on Claude's free plan but is 'exclusive to your system'. About a 5% chance the install fails on OS/hardware; retry two or three times, then ask Perplexity/ChatGPT for OS-specific steps.

Context7 is the demo and 'always' his recommendation before any vibe-coding: it feeds the model current library documentation so generated code stops using stale APIs - he tells Claude to use Context7 'and not skills' when building a sample site. Roughly 30 more servers this way.

Why it matters

Context7 is the single most useful server for anyone generating code with an LLM - though Paul's Catalyst Session 14 later notes it being retired/replaced, so check status.

This idea elsewherecontrastCatalyst Session 14 reports Context7's status changing - verify before recommending
Go deeper

In one line: Medium mode = local Claude Desktop + Node.js; Desktop Extensions tab; free plan OK; device-bound; Context7 for live library docs.

Requires Claude Desktop + Node.js; free plan works (l3186027 1:08-1:11)

Node.js = 'translator' telling Claude where servers live (l3186027 1:12)

Desktop Extensions tab, distinct from the Web tab (l3186027 1:18-1:19)

Context7 recommended first for coding accuracy (l3186027 1:20-1:22)

~5% install failure; retry, then AI-assisted troubleshooting (l3186027 1:13-1:14)

▶ Watch this taught:

04

Stacking servers: one prompt, several MCPs, in a stated order

how-to

One server is a party trick. Two servers in sequence is a workflow you would have paid a developer for.

The power 'is not just learning what MCP servers are' but chaining them: the output of one becomes the input of the next inside a single prompt that spells out the order. Demo: File System MCP server (permissions granted explicitly - Desktop, Downloads, and an added 'AI projects' folder) reads ds_salaries.csv -> Claude does the analysis natively -> Context7 pulls the charting library (recharts) docs -> Claude writes a results website. Proposed stacks: Trivago -> Canva (itinerary to brochure), Trivago -> Context7 (itinerary to website), Google Drive -> Perplexity/Claude web (old report refreshed with current facts).

The rule: 'Explicitly mention the MCP server flow in the prompt' - otherwise, with many servers connected, the model may call the wrong or an unconfigured one.

Do it in this order
Why it matters

This is the mental model behind every 'agent that reads my files and builds me a dashboard' request - Paul's KB itself was built this way.

Go deeper

In one line: Stacking = sequencing several MCP servers in one prompt with the flow stated; configure each server's permissions first.

Chain server outputs to inputs within one prompt (l3186027 1:28-1:30)

File System server needs explicit directory permissions via Add Directory (l3186027 1:31-1:32)

Demo stack: File System -> Claude analysis -> Context7 -> website (l3186027 1:32-1:37)

Other stacks: Trivago->Canva, Trivago->Context7, Drive->Perplexity (l3186027 1:40-1:43)

State the server order in the prompt or the model guesses (l3186027 1:43-1:44)

▶ Watch this taught:

05

Hard mode: Smithery.ai and how to vet a community server

Three thousand seven hundred servers. Four numbers tell you which ones are safe to bolt on.

Smithery.ai (Google or GitHub sign-in) listed 3,721 servers. Vetting: quality score above 60/100 is 'decent good'; monthly tool calls above ~1,500 means real use; uptime above 60%; a recent 'last deployed' date. Two install paths: paste the connector URL into Claude Desktop's Add Custom Connector, or copy the npx command into Terminal/PowerShell, run it, restart Claude. Demoed: Bright Data (web scraping/crawling; ~79/100, ~75% uptime) and Sequential Thinking from the reference servers (70/100, ~6,500 calls/month), stacked with Context7 to draft landing-page copy and then build the page. 'If you are in the beginning stages, stay with easy and medium.'

Why it matters

The vetting numbers are the reusable part - they transfer to any MCP registry, and Paul already has Bright Data as a plugin.

This idea elsewhereparallelBasecamp 5 surveys the same marketplaces
Go deeper

In one line: Hard mode = third-party registries (Smithery) with thousands of servers; vet by quality >60, calls >1,500/mo, uptime >60%, recent deploy; install by URL or npx.

Smithery.ai: 3,700+ servers, Google/GitHub sign-up (l3186027 1:46-1:47)

Vet: quality >60/100, >1,500 monthly calls, uptime >60%, recent deploy (l3186027 1:48-1:50)

Install by custom connector URL or npx command + restart (l3186027 1:51-1:57)

Demoed Bright Data and Sequential Thinking, stacked with Context7 (l3186027 1:50-2:03)

Beginners: stay on easy/medium first (l3186027 1:53, 2:02)

▶ Watch this taught:

06

RAG is memory, MCP is hands - and how to structure messy data for RAG

One remembers. The other does. You usually want both.

'RAG is retrieval augmented generation. It's the memory' - searching and recalling stored knowledge. MCP 'is executing functions and doing things on your behalf that's not solely just pulling information' - GitHub, arXiv, Slack, Gmail servers act on live systems. n8n now has a native MCP node (paste the server JSON) where it used to need a custom node; Gmail MCP is Harshith's favourite example. Adoption spiked - 'look at Google Trends' - and the next layer is A2A: agents' assistants talking to each other with no human in between.

Making RAG good: an unstructured multi-sheet Excel must be normalized into clean structured text and tagged with metadata (a filing cabinet with labelled folders) before embedding. For huge policy/compliance corpora use hybrid search (vector + full text) alongside semantic search and prefer chunks that appear in both - the 'Venn diagram'. A 25 MB 'file too large' error in Pinecone/n8n is usually model context (use GPT-4.1/4o/o1 over 3.5) or a need to pre-chunk.

Why it matters

Answers the question every learner asked twice this week, and gives the two RAG quality moves (normalize + tag, hybrid + semantic) that most tutorials skip.

This idea elsewhereparallelBasecamp 5's MCP-vs-API framing is the sibling distinction
Go deeper

In one line: RAG = retrieval/memory; MCP = action on external systems; n8n has a native MCP node; better RAG = structured, tagged input + hybrid-and-semantic retrieval overlap.

RAG = memory/retrieval; MCP = executing functions on live systems (l3302271 0:39-0:40)

n8n native MCP node: paste server JSON; Gmail MCP favourite (l3302271 0:20-0:21)

Normalize unstructured data to clean text + metadata tags before embedding (l3302271 0:22-0:24)

Hybrid + semantic retrieval; keep the overlapping chunks (l3302271 0:35-0:37)

'File too large' = context size (use 4.1/4o/o1) or pre-chunk (l3302271 0:13-0:14)

A2A = agents' assistants negotiating directly (l3302271 0:34)

▶ Watch this taught:

07

Resistors, adopters, transformers: the Shopify and Dan Shipper cases

At Shopify the engineers did not adopt first. The sales team did - and doubled their close rate with a tool they built themselves.

Shopify: a CEO mandate, and the surprise that non-technical teams moved fastest - a sales team built a page-load-speed comparison against competitors' stores and reportedly doubled close rates over a 6-8 month rollout. Dan Shipper's company: 7-figure revenue, 6-7 products, '100% AI written code', newest hire an 'AI Ops Expert' - which Dileep maps onto this program's 'AI generalist'. Three archetypes: resistors (restrict AI), adopters (10-20% gains), transformers (redesign work around it). Reid Hoffman paraphrased: you spin up work, and while you are in meetings or asleep it gets done. Roadmap teaser: agent orchestration, safety alignment, context engineering at scale.

Why it matters

The 'non-technical adopt fastest' point is the pitch Paul makes to small-business clients.

Go deeper

In one line: Three adopter archetypes; evidence that mandate + non-technical builders yields the fastest wins; 'AI Ops' as the emerging role.

Shopify: CEO mandate; sales team's speed-comparison tool doubled close rates in 6-8 months (l3186027 2:06-2:08)

Dan Shipper: 7-figure, 100% AI-written code, 'AI Ops Expert' hire (l3186027 2:09-2:10)

Resistors / adopters (+10-20%) / transformers (l3186027 2:11-2:12)

Roadmap: orchestration, safety alignment, context engineering at scale (l3186027 2:14-2:15)

▶ Watch this taught:

Every concept, three clicks deep

The same concepts as a quick reference: the closed row is the glance, open is the study card, and every timestamp jumps into the recording.

01API, webhook, MCP: check, get called, or send the conciergeMCP = open protocol (Anthropic) letting an LLM discover and call external tools/data through servers;›

MCP = open protocol (Anthropic) letting an LLM discover and call external tools/data through servers; a concierge model vs API pull and webhook push; requires an LLM in the loop.

API = you check; webhook = they notify; MCP = concierge does the exchange (l3186027 0:17-0:19)

Model + Context + Protocol; 'universal travel adapter' (l3186027 0:21-0:28)

An LLM must be part of the loop or it is not MCP (l3186027 0:28)

Proposed by Anthropic, open-sourced, adopted by competitors incl. OpenAI (l3186027 0:33; l3302271 0:33)

Gauntlet/infinity-stone analogy: each server adds a power (l3302271 0:21)

02Easy mode: native connectors, paid plans, and the token taxEasy mode = vendor-hosted connectors in ChatGPT/Claude web;›

Easy mode = vendor-hosted connectors in ChatGPT/Claude web; zero install, any device, paid plan required; name the server once you have many.

No installation; web-hosted; any signed-in device (l3186027 0:35)

Paid plans only (~$20/mo) (l3186027 0:36, 1:00)

~30 ChatGPT connectors; Claude path: + -> Connectors -> Browse -> Web tab (l3186027 0:39-0:49)

Connector context burns limits faster - by design (l3186027 1:00-1:03)

Server chosen from stored descriptions; name it explicitly past ~10 servers (l3186027 1:04-1:07)

Easy + medium together ~60+ servers (l3186027 1:27)

03Medium mode: Claude Desktop + Node.js, Desktop Extensions, and why Context7 comes firstMedium mode = local Claude Desktop + Node.js;›

Medium mode = local Claude Desktop + Node.js; Desktop Extensions tab; free plan OK; device-bound; Context7 for live library docs.

Requires Claude Desktop + Node.js; free plan works (l3186027 1:08-1:11)

Node.js = 'translator' telling Claude where servers live (l3186027 1:12)

Desktop Extensions tab, distinct from the Web tab (l3186027 1:18-1:19)

Context7 recommended first for coding accuracy (l3186027 1:20-1:22)

~5% install failure; retry, then AI-assisted troubleshooting (l3186027 1:13-1:14)

04Stacking servers: one prompt, several MCPs, in a stated orderStacking = sequencing several MCP servers in one prompt with the flow stated;›

Stacking = sequencing several MCP servers in one prompt with the flow stated; configure each server's permissions first.

Chain server outputs to inputs within one prompt (l3186027 1:28-1:30)

File System server needs explicit directory permissions via Add Directory (l3186027 1:31-1:32)

Demo stack: File System -> Claude analysis -> Context7 -> website (l3186027 1:32-1:37)

Other stacks: Trivago->Canva, Trivago->Context7, Drive->Perplexity (l3186027 1:40-1:43)

State the server order in the prompt or the model guesses (l3186027 1:43-1:44)

05Hard mode: Smithery.ai and how to vet a community serverHard mode = third-party registries (Smithery) with thousands of servers;›

Hard mode = third-party registries (Smithery) with thousands of servers; vet by quality >60, calls >1,500/mo, uptime >60%, recent deploy; install by URL or npx.

Smithery.ai: 3,700+ servers, Google/GitHub sign-up (l3186027 1:46-1:47)

Vet: quality >60/100, >1,500 monthly calls, uptime >60%, recent deploy (l3186027 1:48-1:50)

Install by custom connector URL or npx command + restart (l3186027 1:51-1:57)

Demoed Bright Data and Sequential Thinking, stacked with Context7 (l3186027 1:50-2:03)

Beginners: stay on easy/medium first (l3186027 1:53, 2:02)

06RAG is memory, MCP is hands - and how to structure messy data for RAGRAG = retrieval/memory;›

RAG = retrieval/memory; MCP = action on external systems; n8n has a native MCP node; better RAG = structured, tagged input + hybrid-and-semantic retrieval overlap.

RAG = memory/retrieval; MCP = executing functions on live systems (l3302271 0:39-0:40)

n8n native MCP node: paste server JSON; Gmail MCP favourite (l3302271 0:20-0:21)

Normalize unstructured data to clean text + metadata tags before embedding (l3302271 0:22-0:24)

Hybrid + semantic retrieval; keep the overlapping chunks (l3302271 0:35-0:37)

'File too large' = context size (use 4.1/4o/o1) or pre-chunk (l3302271 0:13-0:14)

A2A = agents' assistants negotiating directly (l3302271 0:34)

07Resistors, adopters, transformers: the Shopify and Dan Shipper casesThree adopter archetypes;›

Three adopter archetypes; evidence that mandate + non-technical builders yields the fastest wins; 'AI Ops' as the emerging role.

Shopify: CEO mandate; sales team's speed-comparison tool doubled close rates in 6-8 months (l3186027 2:06-2:08)

Dan Shipper: 7-figure, 100% AI-written code, 'AI Ops Expert' hire (l3186027 2:09-2:10)

Resistors / adopters (+10-20%) / transformers (l3186027 2:11-2:12)

Roadmap: orchestration, safety alignment, context engineering at scale (l3186027 2:14-2:15)

Tools referenced

ToolCoverageMomentContext
Claude DesktopdemonstratedMedium/hard mode host
Node.jsdemonstratedRequired runtime; 'translator' analogy
Context7demonstratedLive library docs; first recommendation
Smitherydemonstrated3,700+ server directory; vetting metrics
Bright DatademonstratedScraping MCP server installed from Smithery
ChatGPTdemonstratedConnectors menu; Spotify playlist demo
ClaudedemonstratedWeb connectors: Trivago, Google Drive
n8nexplainedNative MCP node - paste server JSON
PineconeexplainedMulti-document RAG; 25MB error diagnosis
OllamaexplainedLocal GPT-OSS/DeepSeek with n8n cloud
CanvamentionedStacking target (itinerary -> brochure)
PerplexitymentionedStacking target and install helper; Perplexity MCP deferred
FirecrawlmentionedScrape multi-link sites into RAG
InstantlymentionedCold-email sending in the lead-gen case study
AnyMailFindermentionedEmail finder in the lead-gen stack
LinkedIn Sales NavigatormentionedTargeting layer of the lead-gen stack
SunomentionedRecommended AI music tool
OpusClipmentionedLong-form to shorts repurposing
TypefullymentionedSocial scheduling

Action items

    Resources mentioned

    Resources
    • docDay 5 workbook - File System MCP on budget.xlsx
    • docOffice Hour 3 question log (~55 topics)

    Extraction notes

    This page was built from an auto-generated transcript, which garbles product and people's names. Those were corrected silently in everything above and logged here for transparency. The warnings flag claims that were true on the recording day but change fast.

    Transcript corrections applied

    The transcript saysThe trainer actually means
    smithree dot a ISmithery.ai
    Contact 7 / contacts 7Context7
    FiretronFirecrawl
    Sophie UIComfyUI
    Defi / DiffieDify
    ProudSocialunresolved social scheduling tool
    Dimitj / Adalweekend mentor names - unresolved

    True on recording day — verify before relying