MCP for Agencies: How to Connect Claude or ChatGPT to Client Work

MCP lets an AI assistant like Claude or ChatGPT read and act on the tools your team already uses, instead of working from whatever you paste into a chat window. For agencies, that means asking about client work in plain words and getting answers from the real requests, briefs, and files.

This guide covers what MCP is, how to connect an assistant to client requests, which permissions to set, what to ask, and where it falls short.

What is MCP, and why should an agency care?

MCP (Model Context Protocol) is an open standard that gives AI applications a common way to connect to external tools and data. Google Cloud’s overview of MCP describes it as an open standard introduced by Anthropic. AWS Builder Center describes it the same way: a standard route for AI applications to reach tools, resources, and data sources.

Without MCP, an assistant only knows what you paste in. You copy a client email into Claude, ask for a summary, then copy the answer back. Every question starts with a round of copying.

With MCP, the assistant connects to a tool and asks it directly. The tool decides what the assistant can see and do. You ask, “What’s waiting on a client today?” and the assistant checks the source.

The three parts in plain terms

There are three pieces, and you only need to understand them at this level:

• The assistant: Claude, ChatGPT, or another AI app you already use.
• The MCP server: a connector run by the tool that holds your client data. It lists what the assistant may ask for and do.
• The permissions: the rules that decide whose data the assistant can reach and which actions it can take.

The third piece is the one agencies skip. A connection is only as safe as the access behind it.

Why this matters for client work specifically

Agency work is scattered. A brief lives in one thread, the logo in a shared drive, the approval in an email, and the revision list in someone’s head.

MCP for agencies matters when the assistant can reach the place where requests, files, and sign-offs actually live. Then the questions you ask all day, “where is this at?” and “what did they say about the hero image?”, get answered from the record, not from memory.

Other vendors are building the same bridge. AgencyAnalytics positions its MCP feature as a way to connect Claude, ChatGPT, or other AI tools to query client performance data. That is a reporting use case. This guide focuses on the other half of agency life: requests, briefs, files, and approvals.

What can an AI assistant do with client requests over MCP?

Connected to a client portal, an assistant can answer questions about requests, summarize threads, open or update requests, and in some setups do the work itself. What it can do depends on the tool and on your access.

With Dora, the AI account manager and client portal, you can connect Claude or ChatGPT over MCP and ask in plain words what needs you today, what’s waiting on a client, or to open a request. The assistant acts with your access, not with extra powers.

A developer or designer can also connect Claude Code and run a command that reads the brief, the thread, and the client’s files, does the work, and submits it for client review. Hand client requests to Claude Code walks through that flow.

Four jobs that suit it well

Some tasks fit an assistant better than others. These four come up most:

1. Morning triage. “What needs me today?” returns the requests waiting on your team, instead of you opening ten tabs.
2. Catching up on a thread. “Summarize where the homepage redesign stands for Acme” saves you from rereading 40 messages before a call.
3. Opening requests from a conversation. You finish a call, then tell the assistant to open a request for the three changes the client mentioned.
4. Handing production work to a coding assistant. A small, well-briefed request, like a landing page tweak, can go to Claude Code and come back for review.

What it should not do

An assistant can be wrong. Dora’s AI runs on Claude, and AI can misread a thread or miss a detail. Check anything important before it reaches a client.

Decisions about the work stay with people. In Dora’s chat, every change she suggests waits for a person’s click. She only assigns requests or approves work on her own if the company turns those settings on.

How do you connect Claude or ChatGPT to your client portal?

You connect an assistant by adding the tool’s MCP connector in the assistant’s settings, then approving access on a consent screen. The exact clicks differ by assistant, so follow the setup page for your tool.

For Dora, connected AI apps use OAuth 2.1 with PKCE, go through a consent screen, act as the approving person, and can be disconnected in Settings. The same permission checks apply in the app, the API, and over MCP.

A numbered setup checklist

Use this order. It keeps access narrow from the start.

1. Decide who connects. Start with one account manager or producer, not the whole team.
2. Check that person’s role. The assistant acts as them, so the role sets the ceiling. If they shouldn’t see a client, the assistant can’t either.
3. Follow the setup page. Use the guide for Claude or ChatGPT and read the consent screen before approving.
4. Run a read-only test. Ask what’s waiting on a client. Compare the answer to what you see in the portal.
5. Try one small write. Open a single low-stakes request and confirm it appears correctly.
6. Review after a week. Note what saved time, what was wrong, and whether the connection should widen or be disconnected.

Prerequisites

You need a Dora account with at least one client, an account on Claude or ChatGPT that supports connectors, and a team member with the right role. Dora’s first client is free, forever, and each client after that is $10 a month, with unlimited team seats and every feature included. See Dora pricing for details.

Which permissions should you set before you connect anything?

Give the assistant the same access as the person who connects it, and no more. The safest MCP setup is one where the assistant cannot see anything its user could not already see.

This is the point where agencies get careless. Attentive’s explainer on MCP for marketers frames governed access to approved data as a prerequisite, not an afterthought. That applies to client work as much as marketing data.

Point of view: narrow first, widen later

Start with one person, one or two clients, and read-heavy questions. Widen only after the answers have been right for a while.

The contrarian position: don’t connect the assistant through a shared admin login so “everyone can use it.” It is faster to set up, but you lose the link between who asked and what was done. Connect each person separately so the assistant acts as that person and the access follows their role.

What Dora’s access model looks like

These are the facts that matter for a security review:

• Each client sees only their own workspace. Internal notes and team-only updates are hidden from clients.
• Team access follows roles in the company.
• Connected apps act as the approving person and can be disconnected in Settings.
• Sign-in is by emailed link, with no passwords. Links expire after 15 minutes.
• Dora’s AI runs on Claude, made by Anthropic. The content she needs (a request, its thread, related files) is sent to Anthropic to write her answer. We don’t train AI models on customer content, and under its commercial terms Anthropic doesn’t either.

Dora does not claim SOC 2, ISO 27001, HIPAA, or SSO. If your clients require any of those, read Security at Dora before you decide. That is a fair reason to choose a different setup.

Internal notes and client-facing text

One risk is easy to miss. If an assistant drafts a reply, it may draw on internal context you never meant to share. Read every client-facing draft before sending, and keep sensitive internal commentary in team-only notes where clients cannot see it.

What should you ask? Example prompts for client work

Good prompts name the client, the request, and the output you want. Vague prompts get vague answers, and an assistant working from a real thread can still guess when the question is loose.

These are examples of the style, not scripts. Adjust the names to your own clients.

Read-only prompts

Start here. Nothing changes, so a wrong answer costs you a minute.

• “What needs me today?”
• “List everything waiting on a client for Northside Dental, oldest first.”
• “Summarize the latest thread on the pricing page request in five lines.”
• “Which requests are in review and have been for more than three days?”

Prompts that change something

Move to these once the read-only answers have held up.

• “Open a request for Northside Dental: update the opening hours in the footer and on the contact page.”
• “Open a request from this call summary, and list what’s missing from the brief.”

The second prompt has a useful habit built in. Ask what’s missing before the work starts. A brief with no deadline or no target page comes back as a question to the client, not as a guess.

A before and after example

Take a five-person design studio. Before: the account manager starts each morning by opening email, the portal, and a shared drive to find what’s stuck. It takes a chunk of the first hour.

After connecting an assistant to the portal: the account manager asks “What’s waiting on a client?” and “What needs me today?”, then reads the two lists. The expected outcome is a shorter morning check, but this is an example, not a measured result. To measure it yourself, time the morning check for one week before connecting, then one week after, and compare.

How do you review what the assistant does?

Treat the assistant like a new junior account manager: useful, fast, and in need of a check on anything a client will see. The review habit matters more than the prompt.

A simple three-part check works for most teams:

1. Source. Does the answer match the request and thread in the portal?
2. Scope. Did it open or change only what you asked for?
3. Send. Is the wording something you’d put your name on?

Pair it with a clear approval step

AI-assisted work still needs sign-off. Route finished work through one-click approvals so a person at the client confirms it, and set a rule for silence. If nobody replies, the work approves itself after the days you choose. Our client approval process guide covers how to set that up without nagging.

Write briefs the assistant can use

An assistant can only work from what’s in the request. If briefs arrive as three lines in an email, the output will be thin. A brief clients will actually fill in fixes the input, and Dora can turn client emails into requests and ask for what’s missing.

Common mistakes and honest limits of MCP for agencies

The biggest mistakes are connecting too broadly, trusting answers without checking, and expecting the assistant to fix messy input. Most problems trace back to one of these.

Mistakes to avoid

• Connecting everything at once. Start with one person and a few clients.
• Skipping the read-only phase. You learn where the assistant is unreliable before it can change anything.
• Pasting client files into a general chat instead. That bypasses the permission checks MCP is meant to give you.
• Treating a summary as the record. The thread is the record. The summary is a shortcut.
• Letting vague requests through. Fix the brief first, then ask for the work.

Limits worth knowing

MCP is still young, and the market is crowded. Loomal’s roundup of marketing MCP servers shows how many exist and flags safety considerations for agent access. Many tools now offer a connector, and competitors such as ManyRequests, Assembly, ClickUp, and Notion have MCP servers too. A connector existing is not the differentiator. What matters is how the AI works for your clients and your team.

An assistant also won’t replace judgment. It can’t tell you whether a client’s request is really scope creep, or whether to push back on a deadline. It can pull the original scope and the thread so you decide faster.

Dora does not do time tracking, invoicing, contracts, or a CRM. If you need those, pair Dora with tools that do them. For internal planning, a tool like ClickUp may be the better fit.

A simple way to measure whether it’s working

Pick one baseline metric, such as minutes spent on the morning status check or the number of “any update on this?” emails per week. Record it for two weeks. Connect the assistant, then measure again for two weeks. If the number hasn’t moved, narrow the use or disconnect.

If you want to try this with your own clients, you can meet Dora, the AI account manager. Your first client is free, forever, so you can connect Claude or ChatGPT to a real client and see what it does before you commit.

FAQ

What is MCP in simple terms?

MCP is a standard way for an AI app to connect to outside tools and data. Instead of pasting information into a chat, the assistant asks the tool directly, within the access it has been given.

Is it safe to connect Claude or ChatGPT to client data?

It can be, if access is narrow and tied to a person’s role. With Dora, connected apps act as the approving person, go through a consent screen, and can be disconnected in Settings. Check anything important the AI produces, and read Security at Dora for details.

Does Dora make changes on its own through MCP?

No, not by default. In Dora’s chat, every change she suggests waits for a person’s click. She only assigns or approves work on her own if the company turns those settings on.

Do I need a developer to set up MCP for agencies?

Not for Claude or ChatGPT. You add the connector, approve the consent screen, and start asking questions. Claude Code is aimed at people who already work in a coding environment.

What does it cost to connect an assistant to Dora?

There is no separate MCP fee. Dora’s first client is free, forever, and each client after that is $10 a month, with unlimited team seats and every feature included.

References

1. Google Cloud: What is Model Context Protocol (MCP)?
2. AWS Builder Center: MCP Explained
3. AgencyAnalytics MCP
4. Attentive: What Is MCP for Marketers?
5. Loomal: Best Marketing MCP Servers for AI Agents (2026)
6. Why This AI Protocol Is the Future of Automated Marketing
7. Why Ad Tech Should be Watching AI’s Hottest Tech: MCP
8. What is MCP & How It Speeds Up AI Agent Building 100X
9. Agent2Agent vs. MCP: Choosing the Right Protocol for AI Agents
10. Tool Servers MCP for AI Agents Explained | Agent Patterns #3