# MCP Agent Use Cases: 8 AI Agents to Build by Role (2026) > MCP agent use cases by role: sales, marketing, social media, support, data and engineering. The MCP servers worth connecting, their real hosted URLs, and the right model for each job. **Source:** https://mcpplaygroundonline.com/blog/mcp-agent-use-cases **Author:** Nikhil Tiwari **Published:** 2026-09-15 **Updated:** 2026-09-15 **Category:** Guide **Reading time:** 15 min read --- 📖 TL;DR - An **MCP agent** is four things: a model, two or three MCP servers, a system prompt, and a guardrail on writes. - **Sales** → HubSpot or Salesforce + Slack + Calendar. **Marketing** → Ahrefs or Semrush + Exa + Notion. - **Social media** → X and TikTok have official ad MCP servers. Organic posting still needs Zapier or your own server. - **Support** → Stripe + Slack + Linear. **Data** → PostHog or Amplitude + your warehouse. **Engineering** → GitHub + Sentry + Linear. - **Three servers is the ceiling** for most agents. Past that, tool-selection accuracy falls off a cliff. - Model choice is a cost decision, not a quality one. **Haiku 4.5 for routing, Sonnet 5 for reasoning, Opus 5 for long chains.** - Connect every server in a [browser tester](/mcp-test-server) before you wire it to an agent. Half of them expose fewer tools than the docs claim. Table of Contents 1. [Anatomy of an MCP Agent](#anatomy) 2. [How to Pick MCP Servers](#pick-servers) 3. [Sales Agent](#sales-agent) 4. [Marketing Agent](#marketing-agent) 5. [Social Media Agent](#social-media-agent) 6. [Customer Support Agent](#support-agent) 7. [Data and Analytics Agent](#data-agent) 8. [Engineering Agent](#engineering-agent) 9. [Project Ops Agent](#ops-agent) 10. [Finance and RevOps Agent](#finance-agent) 11. [Which Model for Which Agent?](#model-picker) 12. [Test Before You Trust](#test-first) 13. [Five Mistakes to Avoid](#mistakes) 14. [FAQ](#faq) Every team I talk to wants the same thing from MCP. Not a protocol tour — **an agent that does their actual job**. The sales lead wants call prep. The marketer wants a content brief that cites real keyword data. The support lead wants refund triage. These are all the same build. Only the servers change. So this guide is organised the way you actually think about it: **by role**. For each one I list the MCP servers worth connecting, the hosted URLs, a model recommendation with its real cost, and example prompts that work on day one. I have kept every URL genuine. If a service has no official hosted MCP server, I say so instead of inventing one. Skip the theory if you want. Jump to [your role](#sales-agent), copy the server list, and have something running in ten minutes. ## Anatomy of an MCP Agent (Four Parts, No More) Strip away the frameworks and _every MCP agent_ is the same four things. Part What it decides Where teams get it wrong **Model** How well it picks the right tool Paying for a frontier model to do lookups **MCP servers** What the agent can actually reach Connecting eight when three would do **System prompt** The job, the tone, the limits Vague role text with no stop rules **Write guardrail** What it may change without asking Skipped entirely until something breaks **The fourth row is the one people skip.** An agent with a Stripe token can issue refunds. An agent with a CRM token can overwrite deal stages. I put one line in every system prompt I write: _propose writes, never execute them, until I say go_. If the protocol itself is new to you, start with [what the Model Context Protocol actually is](/blog/what-is-model-context-protocol), then come back here. ## How to Pick MCP Servers for Any Agent The instinct is to connect everything. **Resist it.** Every connected server injects its full tool schema into the model's context on every single request. A large server like GitHub runs to roughly 18,000 tokens on its own. Connect six of those and you have burned your context window before the user types a word. Worse, tool-selection accuracy drops as near-identical tool descriptions pile up. I use a **three-server rule**, and each server plays a distinct part: 1. **System of record** — where the truth lives. HubSpot, GitHub, Stripe, your warehouse. 2. **Context source** — what fills the gaps. Exa for the web, Notion for internal docs, Calendar for timing. 3. **Delivery channel** — where the output lands. Slack, Notion, Linear. Record, context, channel. **Almost every useful agent in this guide is that shape.** If you genuinely need more surface area, do not add a fourth server. Use tool search instead, which I cover in [MCP context bloat and tool search](/blog/mcp-context-bloat-tool-search). ## The Sales Agent: CRM Hygiene and Call Prep Sales reps lose hours to two chores. **Researching a prospect before a call, and updating the CRM after it.** Both are pure context assembly. That is exactly what an MCP agent is good at. MCP server Hosted URL What it unlocks **[HubSpot](/mcp-servers/hubspot)** `https://mcp.hubspot.com` Deals, contacts, pipeline stages, notes **[Salesforce](/mcp-servers/salesforce)** Per-org URL from your instance Opportunities, accounts, custom objects **[Clay](/mcp-servers/clay)** `https://mcp.clay.earth/mcp` Enrichment, firmographics, contact finding **[Exa](/mcp-servers/exa)** `https://mcp.exa.ai/mcp` Live web research on the account **Google Calendar** `https://calendarmcp.googleapis.com/mcp/v1` Who you are meeting, and when **[Slack](/mcp-servers/slack)** `https://mcp.slack.com/mcp` Delivery channel for the brief **My recommended build:** Calendar + HubSpot + Exa. The agent reads tomorrow's meetings, pulls each account's CRM history, and researches recent news. That is the [Sales Call Prep template](/templates/sales-call-prep), ready to run. For pipeline hygiene instead, use [Sales Pipeline](/templates/sales-pipeline), which pairs HubSpot with Slack and Linear. **Model:** Claude Sonnet 5 at 8 credits a run. Call prep involves synthesis across three sources, which is where cheaper models start dropping details. Prompts that work immediately: - Prep me for tomorrow's calls. For each one, pull the deal stage, last contact date, and anything newsworthy about the company. - Which deals have had no activity in 21 days and are still marked as open? - Draft a follow-up email for the Acme deal referencing what we discussed last time. **Watch the write scope.** A HubSpot token with write access lets the agent change deal stages. Start read-only, and add writes once you trust its judgement. ## The Marketing Agent: Briefs Backed by Real Data Most AI content workflows fail for one reason. **The model has no idea what people actually search for.** It guesses at keywords, invents volumes, and produces a brief that reads well and ranks nowhere. Connecting a real SEO data source fixes that in one step. MCP server Hosted URL Best for **[Ahrefs](/mcp-servers/ahrefs)** `https://api.ahrefs.com/mcp/mcp` Backlinks, keyword difficulty **[Semrush](/mcp-servers/semrush)** `https://mcp.semrush.com/v1/mcp` Competitor gaps, position tracking **[DataForSEO](/mcp-servers/dataforseo)** `https://mcp.dataforseo.com/mcp` Raw SERP data, cheapest per call **[OpenSEO](/mcp-servers/openseo)** `https://app.openseo.so/mcp` AI-search visibility tracking **[Firecrawl](/mcp-servers/firecrawl)** `https://mcp.firecrawl.dev/v2/mcp` Scraping competitor pages to markdown **[Notion](/mcp-servers/notion)** `https://mcp.notion.com/mcp` Where the finished brief lands **[Webflow](/mcp-servers/webflow)** / **[Sanity](/mcp-servers/sanity)** `https://mcp.webflow.com/mcp` `https://mcp.sanity.io` Publishing straight to the CMS **[Canva](/mcp-servers/canva)** `https://mcp.canva.com/mcp` Generating on-brand visuals **My recommended build:** Ahrefs + Exa + Notion. Keyword truth, live SERP context, and a place to file the output. That combination ships as the [SEO Content Strategist template](/templates/seo-content-strategist). If your focus is visibility inside AI answers rather than blue links, use [OpenSEO Research Agent](/templates/openseo-research-agent) instead. **Model:** Sonnet 5, or Sonnet 4.6 at 12 credits when the brief spans a dozen competitor pages. Prompts worth stealing: - Find five keywords we rank on page two for, with difficulty under 30, and outline a post for the best one. - Compare the top five results for this keyword and list every heading they cover that we do not. - Which of our published posts have lost positions in the last 90 days? Before you hand an SEO server real API credits, [connect it in the browser tester](/mcp-test-server) and read its tool list. Ahrefs and Semrush both meter by call, and a chatty agent burns quota fast. ## The Social Media Agent: Where MCP Helps and Where It Does Not This is the role with the biggest gap between expectation and reality, so let me be blunt about it. **Paid social has excellent official MCP coverage. Organic posting mostly does not.** X and TikTok both ship real hosted MCP servers, and both are aimed at advertising. There is no official hosted MCP server for scheduling an Instagram carousel or a LinkedIn post. Platform MCP URL Scope **[X (Twitter)](/mcp-servers/x-twitter)** `https://api.x.com/mcp` Official, ads and API surface **[TikTok for Business](/mcp-servers/tiktok)** `https://business-api.tiktok.com/open_mcp/tt-ads-mcp-flat` Official, campaign management **Meta Ads (Pipeboard)** `https://meta-ads.mcp.pipeboard.co/` Third-party, Facebook and Instagram ads **[Apify](/mcp-servers/apify)** `https://mcp.apify.com` Scraping public profiles and competitor feeds **[Canva](/mcp-servers/canva)** `https://mcp.canva.com/mcp` Generating post creative from a brand template **[Zapier](/mcp-servers/zapier)** Per-account URL you generate The practical bridge to organic posting **Zapier is the honest answer for scheduling.** You pick the actions you want exposed, Zapier generates a private MCP endpoint, and the agent calls those actions as tools. It is not elegant. It works today, and it covers the platforms nobody else does. For paid social, the [Meta Ads template](/templates/meta-ads-pipeboard) gives you spend, ROAS and creative performance in one conversation. **Model:** Claude Sonnet 5 for ad analysis. Haiku 4.5 at 4 credits is plenty if the agent only drafts copy and reads metrics. Prompts to start with: - Which three ad sets had the worst ROAS last week, and what do the winning ones have in common? - Pull our last 20 posts and tell me which format gets the most saves. - Draft five hooks for this week's launch, each under 200 characters, in our usual voice. ## The Customer Support Agent: Triage Before a Human Reads It Support tickets arrive with no context. **The agent's job is to attach it before a human opens the thread.** Who is this customer? What plan are they on? Did they just get charged twice? Is this a known bug? Three servers answer all four questions. - **[Stripe](/mcp-servers/stripe)** — `https://mcp.stripe.com/` for subscription state, invoices and failed payments. - **[Slack](/mcp-servers/slack)** — `https://mcp.slack.com/mcp` to read the support channel and post the summary back. - **[Linear](/mcp-servers/linear)** — `https://mcp.linear.app/sse` to check whether the bug is already filed. That exact trio is the [Customer Success template](/templates/customer-success). It turns a one-line complaint into a briefed ticket. **Model:** Haiku 4.5 at 4 credits. Triage is high-volume and low-ambiguity, which is the cheapest model's sweet spot. This is also the role where _read-only really matters_. A support agent should never issue a refund on its own initiative. **Prompt injection is a live risk here.** Ticket text is untrusted input written by strangers. An agent that reads tickets and holds a Stripe write token is one crafted message away from a bad day. Keep the refund tool out of its reach and [scan the servers you connect](/mcp-security-scanner). ## The Data and Analytics Agent: Answers Without a Ticket Every analytics team has the same queue. **Twenty people asking questions that are one SQL query away from an answer.** An MCP agent with warehouse access clears most of that queue, as long as you keep it read-only. Product analytics [PostHog](/mcp-servers/posthog) `https://mcp.posthog.com/mcp` [Amplitude](/mcp-servers/amplitude) `https://mcp.amplitude.com/mcp` Warehouses BigQuery `https://bigquery.googleapis.com/mcp` [Snowflake](/mcp-servers/snowflake) and [ClickHouse](/mcp-servers/clickhouse) run locally App databases [Neon](/mcp-servers/neon) `https://mcp.neon.tech/sse` [MongoDB](/mcp-servers/mongodb) via its local server Notebooks Hex `https://app.hex.tech/mcp` **My recommended build:** one analytics source, one warehouse, Slack for delivery. The [Data Analyst template](/templates/data-analyst) wires Hex, Neon and Slack together. For product teams, [Product Analytics Ops](/templates/product-analytics-ops) pairs Amplitude with Linear so funnel drops become tickets. **Model:** Sonnet 4.6 at 12 credits. SQL generation against an unfamiliar schema rewards the stronger reasoning models. **Non-negotiable:** connect with a read-only role. Not a role you promise to use carefully — one the database will not let write. ## The Engineering Agent: The Best-Served Role in MCP Engineering has the deepest MCP coverage of any function. Almost every developer tool shipped a server first. Server URL Agent job **[GitHub](/mcp-servers/github)** `https://api.githubcopilot.com/mcp/` PR review, release notes, stale-branch sweeps **[Sentry](/mcp-servers/sentry)** `https://mcp.sentry.dev/mcp` Error triage, regression spotting **[Linear](/mcp-servers/linear)** `https://mcp.linear.app/sse` Sprint state, ticket creation **[Vercel](/mcp-servers/vercel)** `https://mcp.vercel.com` Deploy status, build log reading **[Cloudflare](/mcp-servers/cloudflare)** `https://mcp.cloudflare.com/mcp` Workers, DNS, edge config **[Figma](/mcp-servers/figma)** `https://mcp.figma.com/mcp` Design-to-code handoff **[Context7](/mcp-servers/context7)** `https://mcp.context7.com/mcp` Current library docs, no auth needed **[Datadog](/mcp-servers/datadog)** `https://mcp.datadoghq.com/api/unstable/mcp-server/mcp` Live metrics and logs during an incident **PagerDuty** `https://mcp.pagerduty.com/mcp` On-call context and incident timelines Three builds cover most of what engineering teams ask for: - **[Engineering Lead](/templates/engineering-lead)** — GitHub + Linear + Slack. Sprint status without the standup. - **[Full Stack Shipper](/templates/full-stack-shipper)** — GitHub + Vercel + Sentry. Ship, watch, roll back. - **[Post-Incident Review](/templates/post-incident-review)** — Sentry + Linear + Slack. A timeline written while it is still fresh. **Model:** Sonnet 4.6 or Opus 5 at 15 credits. Code reasoning across a diff is the one place the top tier consistently earns its cost. Only the GitHub server is free to connect with no key. Context7 is the other one, which makes the pair a good first test. ## The Project Ops Agent: Meetings In, Tasks Out Project managers spend their week converting one format into another. **Meeting notes into tickets. Tickets into status updates.** That is mechanical work, and it is the easiest agent on this list to get right. - **[Asana](/mcp-servers/asana)** — `https://mcp.asana.com/mcp` - **[Monday.com](/mcp-servers/monday)** — `https://mcp.monday.com/mcp` - **[Jira and Confluence](/mcp-servers/jira)** — `https://mcp.atlassian.com/v1/mcp` - **Airtable** — `https://mcp.airtable.com/mcp` - **Google Calendar** — `https://calendarmcp.googleapis.com/mcp/v1` **My recommended build:** Calendar + Notion + Slack, which is the [Meeting Follow-Up template](/templates/meeting-follow-up). It reads yesterday's meetings, finds the notes, and drafts the action items. For sprint rituals, [Product Sprint](/templates/product-sprint) combines Linear, GitHub and Slack. **Model:** Haiku 4.5. Summarising and restructuring text does not need a reasoning model, and this agent runs daily. ## The Finance and RevOps Agent: Revenue Questions, Answered Live Finance questions are usually simple and always urgent. **Which subscriptions failed to renew this week?** Payment platforms have solid MCP coverage, so this one is quick to stand up. - **[Stripe](/mcp-servers/stripe)** — `https://mcp.stripe.com/` for subscriptions, invoices, disputes and payouts. - **[PayPal](/mcp-servers/paypal)** — `https://mcp.paypal.com/mcp` for orders and refunds. - **[Cashfree](/mcp-servers/cashfree)** — `https://mcp.cashfree.com/mcp` for India-first payment flows. The [Revenue Ops template](/templates/revenue-ops) joins Stripe to Linear and Slack, so a spike in failed payments becomes a tracked issue rather than a Slack message nobody actions. **Model:** Sonnet 5. Money questions deserve a model that checks its arithmetic against the tool output rather than guessing. ## Which Model for Which Agent? Teams overthink this. **Model choice is a cost decision far more often than a quality one.** A triage agent reading tickets does not need frontier reasoning. A code-review agent across a 900-line diff does. Here is how I map them, with the per-run credit cost from [MCP Agent Studio](/mcp-agent-studio): Model Credits Use it for **GPT-5.4 nano** 2 One-tool lookups, routing, Zapier actions **Claude Haiku 4.5** 4 Support triage, standups, daily digests **Claude Sonnet 5** 8 The default. Sales, marketing, finance **Claude Sonnet 4.6** 12 Three servers, long tool chains, SQL **Claude Opus 5** 15 Code review, incident analysis, hard chains **Gemini 3 Flash** 2 Web research with a big result set **Start one tier below what you think you need.** If the agent picks the wrong tool or drops a step, move up one and compare. That comparison is worth running properly rather than by vibes. You can put the same prompt through several models side by side on the [MCP model comparison page](/mcp-model-comparison). If you want the reasoning behind the tiers in more depth, I wrote it up in [the best AI model for MCP tool calling](/blog/best-ai-model-for-mcp-tool-calling). ## Test Before You Trust: Four Checks That Take Five Minutes Here is the failure mode I see most. Someone wires four servers into an agent, it behaves strangely, and they blame the model. Nine times out of ten the server was the problem. **It exposed three tools instead of the twelve the docs promised**, or its auth silently failed. So before any URL in this guide goes into an agent, run it through four checks: 1. **Does it connect at all?** Paste the URL into [the MCP server tester](/mcp-test-server) and watch the handshake. 2. **What tools does it really expose?** Read the list. Compare it to what you assumed. 3. **Does a real call return real data?** A tools list that loads is not proof the tool works. Invoke one. 4. **How big is the schema?** A server with 40 verbose tools will crowd out everything else you connect. **This runs in the browser with no install.** I do it for every server before it touches a production agent, and it has saved me a lot of confused debugging. **MCP Playground is built for exactly this step.** Paste any remote MCP URL, complete the OAuth flow or add a bearer token, and inspect the tool list and raw JSON-RPC responses before you commit. [Test any MCP server free →](/mcp-test-server) Then scan it. A third-party server you found in a directory gets a full audit before it holds a token — [scan your MCP server →](/mcp-security-scanner) ## Five Mistakes That Ruin Otherwise Good Agents **1\. Connecting every server you can find.** Three is the working ceiling. Each extra one costs context and accuracy. **2\. Giving write access on day one.** Run read-only for a week. Read the transcripts. Then decide which writes it has earned. **3\. Using one generic system prompt for every agent.** A sales agent and a support agent need different stop rules, not the same helpful-assistant boilerplate. **4\. Skipping the token scope review.** A GitHub token with repo scope reaches every private repository you can. Scope it down before it goes in. **5\. Never evaluating the thing.** If you cannot say whether last week's version was better, you are guessing. [Evals for MCP agents](/blog/mcp-eval-engine-design) covers how to set that up. ## Frequently Asked Questions **How many MCP servers should one agent connect?+** Two or three for almost every role in this guide. Each connected server injects its full tool schema into context on every request, and a large one like GitHub runs to roughly 18,000 tokens by itself. Past three servers, tool descriptions start overlapping and the model picks wrong more often. If you genuinely need a wider surface, enable tool search rather than adding a fourth server. **Is there an official MCP server for posting to LinkedIn or Instagram?+** No. Paid social is well covered, with official hosted servers from X at api.x.com/mcp and TikTok for Business, plus third-party Meta ads servers. Organic scheduling is the gap. The practical route is a Zapier MCP endpoint, where you choose which actions to expose and Zapier generates a private URL, or writing a small server against the platform API yourself. **Which model should a sales or marketing agent use?+** Claude Sonnet 5 is the right default for both, at 8 credits per run in Agent Studio. Both roles synthesise across two or three sources, which is where cheaper models start dropping details. Drop to Haiku 4.5 if the agent only summarises or drafts copy from a single server, and move up to Sonnet 4.6 when a single run touches a dozen competitor pages or a long CRM history. **Do I need to write code to build an MCP agent?+** Not for the builds in this guide. Every one of them is a model, two or three server URLs, and a system prompt, which you can assemble in Agent Studio or any MCP-capable client. Code becomes necessary when you need a server that does not exist yet, such as organic social posting or an internal system, or when you want the agent to run on a schedule rather than in a chat window. **How do I stop an agent from changing data it should not touch?+** Control it at the token, not in the prompt. Issue a read-only credential wherever the platform supports one, such as a read-only database role or a scoped API key, so a write is impossible rather than merely discouraged. Add a system prompt rule that the agent proposes writes and waits for confirmation, and treat that as a second layer rather than the only one. **Can one agent serve several roles at once?+** It is usually worse than two focused agents. A combined sales and support agent needs both CRM and billing servers connected at all times, which doubles the context cost of every request and blurs the system prompt. Separate agents stay cheaper, pick tools more accurately, and let you give each one a different write scope. ## Start With One Role, Not a Platform The pattern repeats across every role here. **One system of record, one context source, one delivery channel, and a model matched to the difficulty.** Sales gets a CRM and a calendar. Marketing gets keyword data. Support gets billing state. Engineering gets the repo and the error tracker. Pick the role that loses the most hours this week and build that one. **Test every server in the browser before you trust it with a token.** **Check any server before it reaches your agent** Paste a URL, complete auth, and inspect every tool and raw response in your browser. No install, no sign-up. [Test any MCP server free →](/mcp-test-server) [Open Agent Studio →](/mcp-agent-studio) ## Frequently asked questions ### How many MCP servers should one agent connect? Two or three for almost every role. Each connected server injects its full tool schema into context on every request, and a large one like GitHub runs to roughly 18,000 tokens by itself. Past three servers, tool descriptions start overlapping and the model picks wrong more often. If you genuinely need a wider surface, enable tool search rather than adding a fourth server. ### Is there an official MCP server for posting to LinkedIn or Instagram? No. Paid social is well covered, with official hosted servers from X at api.x.com/mcp and TikTok for Business, plus third-party Meta ads servers. Organic scheduling is the gap. The practical route is a Zapier MCP endpoint, where you choose which actions to expose and Zapier generates a private URL, or writing a small server against the platform API yourself. ### Which model should a sales or marketing agent use? Claude Sonnet 5 is the right default for both, at 8 credits per run in Agent Studio. Both roles synthesise across two or three sources, which is where cheaper models start dropping details. Drop to Haiku 4.5 if the agent only summarises or drafts copy from a single server, and move up to Sonnet 4.6 when a single run touches a dozen competitor pages or a long CRM history. ### Do I need to write code to build an MCP agent? Not for these builds. Every one of them is a model, two or three server URLs, and a system prompt, which you can assemble in Agent Studio or any MCP-capable client. Code becomes necessary when you need a server that does not exist yet, such as organic social posting or an internal system, or when you want the agent to run on a schedule rather than in a chat window. ### How do I stop an agent from changing data it should not touch? Control it at the token, not in the prompt. Issue a read-only credential wherever the platform supports one, such as a read-only database role or a scoped API key, so a write is impossible rather than merely discouraged. Add a system prompt rule that the agent proposes writes and waits for confirmation, and treat that as a second layer rather than the only one. ### Can one agent serve several roles at once? It is usually worse than two focused agents. A combined sales and support agent needs both CRM and billing servers connected at all times, which doubles the context cost of every request and blurs the system prompt. Separate agents stay cheaper, pick tools more accurately, and let you give each one a different write scope. --- _Canonical page: https://mcpplaygroundonline.com/blog/mcp-agent-use-cases — MCP Playground (mcpplaygroundonline.com), the free browser-based tool for testing MCP servers and building AI agents._