Claude MCP & ChatGPT MCP as Your AI Agent Builder: Build, Test and Audit Customer-Facing AI Agents

Connect Convocore MCP in three steps

Convocore runs a remote MCP server, so there is nothing to install and no key to copy. Add this URL to Claude, ChatGPT, Cursor or any app that supports remote MCP servers: https://mcp.convocore.app/mcp

Your AI app finds the sign-in on its own. Convocore's server answers with a standard OAuth sign-in and the app registers itself, so you never paste a secret. You sign in with your Convocore account, choose a workspace, click Allow, and you are connected. MCP is included on Convocore's Business plan and above.

Step by step in your AI app

Step one is a different screen in each app. Step two is the same Convocore sign-in page everywhere.

• The fastest path is inside Convocore: the Try out Convocore MCP card in the left sidebar has buttons for Claude, ChatGPT and Cursor that prefill the name and URL for you.
• For a client that cannot do OAuth, such as a script, send your workspace API secret instead: claude mcp add --transport http convocore https://mcp.convocore.app/mcp --header "Authorization: Bearer YOUR_WORKSPACE_SECRET" --header "X-Convocore-Region: na-gcp" (use eu-gcp for an EU workspace). Create the secret under Account, API Secrets.

What you will see when you sign in

The sign-in page opens on convocore.app, not inside your AI app. If you are signed out it asks you to log in first and then brings you back to the same screen. It then asks whether to allow your AI app to use Convocore, and lists the workspaces you administer, in both the EU and NA regions.

• Allow gives that app access to manage the agents, conversations, knowledge bases and settings of the workspace you picked. It creates an API secret named MCP: followed by the app name in that workspace.
• To disconnect, revoke that secret in Convocore under Account, API Secrets. The app loses access straight away.
• The page also shows the address access will be sent to. For Convocore's own server that is mcp.convocore.app.

Check it works, and what to ask first

Open a new chat and make sure the Convocore connector is switched on for it (in Claude, press + then Connectors). Then ask something that needs Convocore.

• Check the connection: List my Convocore agents.
• Build an agent: Scrape https://yourbusiness.com and create a customer service agent for it with a knowledge base, booking and human handoff. Then test it as a customer and fix anything that fails.
• Audit an agent: Read last week's conversations for my agent, find where it gave wrong answers and fix its prompt.
• A workspace is greyed out on the sign-in page: you are not an admin of it. Ask an admin to connect it or to make you one.
• No workspace is listed: create one in the Convocore dashboard first.
• Connected to the wrong workspace: remove the connector in your app, revoke the MCP secret, then add it again and pick the right workspace.
• On the Free, Starter and Pro plans the install buttons inside Convocore are locked until you upgrade to Business.
• Still stuck: remove the connector, add it again from scratch, or write to support@convocore.ai.

What MCP is, and how fast it spread

Anthropic introduced the Model Context Protocol on 25 November 2024 as an open standard for connecting AI assistants to the systems where data lives, so that an assistant can read from and act in other software through a common interface instead of one custom integration per app ¹.

Adoption moved quickly. By April 2025 OpenAI had added MCP support to its Agents SDK and said the ChatGPT desktop app and Responses API would follow ². On 11 September 2025 ChatGPT's Developer Mode gave Plus and Pro users full MCP access, including tools that write and change data, not only read it ³. On 9 December 2025 Anthropic donated MCP to the Agentic AI Foundation under the Linux Foundation, co-founded with Block and OpenAI, and reported more than 10,000 active public MCP servers, 97 million monthly SDK downloads and more than 75 connectors in Claude's directory ⁴.

For anyone building AI agents, the practical change is that the AI app you already use can now do the work inside the platform where your agents live, instead of telling you which buttons to press.

Where customers actually talk to AI agents

A customer-facing agent is not one chat box. In 285,397 real business conversations on Convocore between January and March 2026, website chat carried 55.5%, other chat surfaces 18.4%, and Messenger, WhatsApp and Instagram together about 23%. Voice calls were 0.6% of that snapshot ⁹.

The conversations are short. The median thread was 3 messages, nearly half ended within 2 or 3 messages, and only about 12% ran to 10 messages or more ⁹. That means an agent has very few turns to get the answer right, and a wrong first answer is often the only answer the customer sees.

WhatsApp alone passed 3 billion monthly users in 2025, according to reports of Meta's first quarter results ¹¹, which is why the same agent usually needs to answer on the website and in messaging apps at the same time.

Why every agent has to be tested before customers see it

The τ-bench study from Sierra put AI agents in simulated customer service conversations for a retailer and an airline, with real policies and tools. Even state-of-the-art function-calling agents such as GPT-4o succeeded on fewer than half of the tasks, and they were inconsistent: in retail, fewer than 25% of tasks were solved in all 8 of 8 repeated attempts ⁶. An agent that works when you try it once can still fail the next customer who asks the same thing.

We saw the same pattern in our own test. We ran 9 models across 4 real hotel and venue agents, 3 questions each, for 108 live turns with the real prompts and knowledge bases. Average scores ranged from 6.83 to 10 out of 10, and the share of answers flagged as hallucinations ranged from 0% to 42% depending on the model. One model invented a 50 dollar exotic animal fee; another invented a 13th floor pool ¹⁰.

Anthropic's own guidance on building agents says the same thing in general terms: it recommends extensive testing in sandboxed environments with appropriate guardrails, and says the key to success is measuring performance and iterating ⁷. Testing used to be the slow, manual part. With MCP, the assistant that built the agent can also chat with it as a customer, score the answers and fix the instructions.

What an AI agent developer does with Convocore MCP

Convocore's MCP server gives Claude, ChatGPT, Cursor and Claude Code 146 tools across a Convocore workspace, served from a remote server at https://mcp.convocore.app/mcp, so there is nothing to install. The assistant chooses the tools; you describe the result you want.

In practice it runs the same loop a human agent developer would, only in minutes: scrape the business's website for real services, prices, hours and photos; write the agent's instructions from those facts; create the agent with lead capture, booking and human handoff; fill the knowledge base from a sitemap or a list of pages; chat with the agent as a customer over many turns and run automatic test suites on its answers, knowledge and tools; read real conversations to find where it failed; and fix the instructions and test again. It backs up an agent before large edits and can roll back in one step.

Then it ships the agent: a no-login link to connect WhatsApp, Instagram or Messenger, the website embed code, phone numbers and outbound call campaigns for voice, and connections to Google Calendar, Outlook, Google Sheets, Shopify, Calendly, Zoho CRM and others.

Keeping it safe

Giving an AI write access to a live customer-facing agent needs the same care Anthropic recommends for any agent: testing in a sandbox before going live, and checkpoints where a person can approve ⁷. In Convocore that means testing on a copy or a prototype link before switching a channel on, taking a backup before large edits, and leaving your AI app's approval prompts on for tools that change things.

Use a separate API secret for MCP so you can revoke it without touching anything else. Agencies can give each white label client an AI connection that sees only that client's agents, under the agency's brand, and turn it off at any time.

What the evidence does not show

• MCP adoption figures (servers, downloads, connectors) are Anthropic's own and were published by the company that created the protocol ⁴. We found no independent count.
• τ-bench tested models from mid 2024 in two simulated domains; newer models do better, and its authors now maintain an updated version ⁶. It shows why consistency must be tested, not how often today's agents fail.
• Our hallucination test covered 108 turns on 4 hospitality agents and used an AI judge. It is too small to rank models and is not about MCP itself ¹⁰.
• Our channel data covers Convocore customers in one quarter and reflects which channels they turned on ⁹.
• Gartner's 80% figure is a forecast, not a measurement, and we read it through a press report because Gartner's page blocked us ⁸.
• We have no controlled measurement yet of how much time an AI agent developer saves compared with building by hand; the comparison above is a description of the workflow, not a result.

If you build AI agents for a living

The evidence above points one way: customers are spread across chat and messaging channels, conversations are short, and agents that seem to work can still fail on the next try ^6,9,10. The work that decides whether an agent is good is the testing and auditing, and that is exactly the part an AI agent developer connected through MCP can now do for every agent you ship.

Agencies on Convocore's White Label plan can do this for every client, and give each client their own AI connection under the agency's brand. We sell that plan, so check the numbers here and decide for yourself.

Method and limits

• Every outside number on this page comes from a source we opened on 29 October 2026, or from a report of it where a note says so.
• Sources are labelled primary (read at the original publisher), secondary (a report of the original), commissioned or forecast. Convocore's own data is labelled as ours.
• Charts redraw numbers already given in the text, with the same citations. Where we rounded, the source gives the exact figure.
• Convocore sells the MCP server and the AI agent platform described here, so it has a commercial interest. If you find an error, write to support@convocore.ai and we will correct it.

Sources

1. Introducing the Model Context ProtocolPrimary source
   Anthropic. 25 November 2024.
2. OpenAI adopts Model Context Protocol to boost LLM tool integrationSecondary report
   DeepLearning.AI, The Batch. 16 April 2025.
   A report of OpenAI's announcement. We give the report's date, not the date OpenAI first announced it.
3. OpenAI has launched developer mode for ChatGPT with full access to Model Context ProtocolSecondary report
   The Decoder. 11 September 2025.
   A report of the launch. Developer Mode was for Plus and Pro users on the web at the time.
4. Donating the Model Context Protocol and establishing the Agentic AI FoundationPrimary source
   Anthropic. 9 December 2025.
   Source of the 10,000+ servers, 97M+ monthly SDK downloads and 75+ Claude connectors figures. Anthropic created MCP.
5. Submit a connector to the directoryPrimary source
   Claude documentation, Anthropic. Read 29 October 2026.
   Requirements for remote MCP connectors in Claude's directory: HTTPS, OAuth 2.0, and a title and read or write annotation on every tool.
6. τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains (Yao, Shinn, Razavi, Narasimhan)Primary source
   arXiv 2406.12045. 17 June 2024.
   We read the abstract. The authors work at Sierra, which sells customer service agents. The repository now marks the original retail and airline tasks as outdated.
7. Building effective agentsPrimary source
   Anthropic. 19 December 2024.
   Anthropic sells the models discussed and notes that some of the tooling advice has since changed.
8. Gartner predicts agentic AI will autonomously resolve 80% of common customer service issues without human intervention by 2029Forecast
   CMSWire, reporting Gartner. 5 March 2025.
   A forecast. Gartner's own page blocked our access, so we read the release as republished by CMSWire.
9. What Businesses Actually Use AI Agents For in 2026 (285,397 conversations, Q1 2026)Primary source
   Convocore. 12 August 2026.
   Convocore's own platform data. We sell the platform the conversations ran on.
10. The Best Value LLMs for Text AI Agents in 2026 (108 live turns on real agents)Primary source
    Convocore. 12 August 2026.
    Our own test: 9 models, 4 hotel and venue agents, 3 questions each, scored by an AI judge.
11. WhatsApp crosses 3 billion monthly users as Meta doubles down on AI strategySecondary report
    Nairametrics, reporting Meta's Q1 2025 earnings call. 1 May 2025.
    A report of Mark Zuckerberg's remarks on the earnings call. We did not read Meta's transcript.
12. Add a connector that isn't in the directoryPrimary source
    Anthropic. Read 30 October 2026.
    Where to add a custom connector in claude.ai, and which plans can.
13. Connect Claude Code to tools via MCPPrimary source
    Anthropic. Read 30 October 2026.
    The claude mcp add command and the /mcp Authenticate step.

Questions

Related