Connect Your AI Agent to a Remote MCP Server

Quickchat AI published a step-by-step guide showing how to connect an AI agent to a remote MCP server, using its free plan and a catalog of over 240 servers. The example connects a support agent for Hoist to DeepWiki, a public MCP server from Cognition, to answer questions about the tus-js-client library from its live GitHub source. The guide details the three-step flow: the model picks a tool, Quickchat AI enforces whitelist and auth, and the server runs the tool, with the exchange recorded in the Inbox.

Your AI Agent answers from the knowledge you give it. Some questions need information that does not fit in a knowledge base: something current, something computed, or something that lives in a system you do not own. A remote MCP server is how your Agent reaches out for that information in the middle of a conversation. This guide connects a Quickchat AI Agent to one, with no code, on the free plan: you will pick a server from a catalog of more than 240, switch on exactly one of its tools, steer it with one prompt paragraph, and watch the real call happen in your Inbox. We build it for real. The example uses Hoist , a resumable file-upload API for developers, and DeepWiki , a public MCP server from Cognition that answers questions about any public GitHub repository. Every setting, URL, and prompt below is the exact one used to produce the screenshots. If you already run an agent, the same steps connect it to any of the live servers in the table below which-remote-mcp-server-should-you-connect-first , and if you would rather delegate the clicking, an AI assistant can build this whole setup for you have-chatgpt-or-claude-build-this-for-you over Quickchat AI’s own MCP server. This is the hands-on half of a pair. If you want the protocol background first, read what MCP is and how it works https://quickchat.ai/post/mcp-explained ; this guide spends its words on doing. What you will build A support Agent for Hoist that does two things. It answers Hoist’s own questions plans, file-size limits, storage regions, webhooks from its knowledge base. And when a visitor asks how the open-source tus-js-client library behaves, the library Hoist’s customers install in their own apps, it calls a remote MCP tool to answer from that library’s live source on GitHub, then relates the answer back to using Hoist. That second case is the point. You cannot and should not copy an entire upstream library’s internals into your knowledge base, but your Agent can look them up on demand. You need two things: a free Quickchat AI account sign up here https://app.quickchat.ai/register and about fifteen minutes. The server URL, the tool whitelist, and the prompt are all in this guide. How does an AI agent connect to a remote MCP server? A remote MCP action points your Agent at a server, discovers the tools that server offers, and lets you choose which of them the Agent may call. MCP, the Model Context Protocol, is a standard way for a server to expose a set of tools to an AI. The server publishes each tool’s name, its description, and the arguments it takes; any MCP client can list those tools and call them. When ChatGPT or Claude offer the same servers as connectors in their settings, this is the machinery behind the word. That is the whole trick, and it is why one connection screen works for a scheduling server, a payments server, and a documentation server alike. Three things happen when your Agent uses a remote MCP tool. The three-step flow: the model picks a tool, Quickchat AI enforces your whitelist and auth, the server runs it. The whole exchange lands in your Inbox. Your Agent, the model, picks a tool. It sees every tool you enabled, each with the description the server wrote for it, and decides when to call one, the same way it decides to use any AI Action. Quickchat AI enforces your rules. Before anything is sent, it checks your tool whitelist, attaches your connection headers or OAuth token if the server needs them, and refuses any server address that resolves to an internal network. The server runs the tool and returns data. The Agent uses the result in its reply, and the whole exchange is recorded in the Inbox. Two kinds of value go into a tool call, and telling them apart matters later. The ones the integration fills in are deterministic: the server URL, headers, and tokens, which the model cannot get wrong. The ones the model fills in are judgment calls: the arguments each tool takes, such as a search query or a repository name, which it can get wrong. Everything in the tuning section comes back to this split. What changed in MCP in 2026 If you last read about MCP when every tutorial mentioned SSE, the ground has shifted. The 2026-07-28 revision of the spec https://blog.modelcontextprotocol.io/posts/2026-07-28/ made the protocol core stateless, removed the session header from streamable HTTP, and deprecated the old HTTP-plus-SSE transport with a migration window the changelog https://modelcontextprotocol.io/specification/2026-07-28/changelog is short and worth a skim, and Claude shipped support the same week https://claude.com/blog/bringing-mcp-2026-07-28-to-claude . The practical consequence for you is pleasant: a remote MCP server is now just a plain HTTPS endpoint, no different to operate against than any other web service. Quickchat AI speaks streamable HTTP, and still falls back to the legacy SSE transport when a server’s URL path ends in /sse , so older servers keep working. Three revisions took remote MCP from an SSE pair with sessions to one stateless HTTPS endpoint. The bracket is HTTP plus SSE’s 12-month runway. Step 1: Create the Agent and give it a knowledge base Start with a working support Agent, so there is something of its own to answer from before you add a live tool. Create a free account and a new AI Agent, and give it a name ours is Hoist . Under Identity , set its job in the AI Main Prompt the full prompt comes in Step 4 . Hoist’s Identity page. The name and the AI Main Prompt are the two fields this guide fills; the Guidelines hold the product facts. Then add a short knowledge base: Hoist’s two plans and their limits, how to get an API key, its storage regions, and its upload.completed webhook. This is the baseline. Everything the Agent can already answer, it should answer from here. The MCP tool is only for the questions this knowledge cannot hold. Which remote MCP server should you connect first? Pick a server whose tools answer questions your knowledge base cannot, and prefer a keyless one for your first connection so you can finish in a single sitting. In your Agent, open Actions & MCPs , click Add Action , and choose MCP . The first thing you see is a catalog: more than 240 remote MCP servers with search and categories, from scheduling and CRM to payments, e-commerce, and developer tools. The MCP catalog: 242 servers behind one search box. Picking a tile starts the connection immediately. Three details in this screen are worth knowing: Every tile is one click. Picking a tile fills in the server’s URL and connects immediately; there is nothing to configure first. A sparkle badge means “Tuned by Quickchat”. For those servers, Quickchat AI ships a recommended tool set: connect Calendly, for example, and only the five tools an agent actually needs for scheduling come enabled out of its several dozen, with sensible defaults you can reset to anytime with Use recommended . The catalog is curated, not exhaustive. Servers that need a per-store or per-instance URL live behind Enter a URL manually at the bottom, which accepts any MCP server on the internet. The catalog leans toward servers your customers would benefit from. For this tutorial we use DeepWiki, which sits in the Search & web category, and because “which server is actually worth connecting” is the question every list post dodges, here is ours, tested rather than recycled. Eight keyless remote MCP servers you can try today Each of these answered a real MCP initialize request on August 28, 2026 . No account, no key: paste the URL, or click its catalog tile, and you are connected. | Server | What its tools do | URL | |---|---|---| DeepWiki | https://mcp.deepwiki.com/mcp Upstash https://upstash.com/blog/context7-mcp Context7 https://mcp.context7.com/mcp Hugging Face https://huggingface.co/mcp Microsoft Learn https://learn.microsoft.com/api/mcp Cloudflare Docs https://docs.mcp.cloudflare.com/mcp GitMCP https://gitmcp.io/{owner}/{repo} Firecrawl https://mcp.firecrawl.dev/mcp Exa https://mcp.exa.ai/mcp Each server name links to its official documentation. Beyond the keyless tier, the account-bound servers you have heard of GitHub, Linear, Notion, Stripe, Sentry, Atlassian, PayPal, Canva are all live too; they authenticate with an OAuth popup instead of a key what-if-the-server-needs-a-sign-in . One warning from testing all of this: several “best MCP servers” roundups circulate endpoints that do not respond, or domains that do not resolve at all, because lists copy other lists. Trust a URL that answers an initialize request, or a catalog that someone maintains, over a listicle. Connect the server in one click For a keyless server, connecting is the shortest step in this guide. Find DeepWiki in the catalog the search box gets you there fastest and click its tile. Search filters the catalog as you type: one result left, one click to connect. Quickchat AI contacts the server, verifies it speaks MCP, and lists the tools it found, all within a few seconds. The probe in progress: Quickchat AI verifies the server speaks MCP and lists its tools, inside a 20 second budget. When the check completes, the action is created and switched on automatically . DeepWiki publishes three tools, and all three arrive enabled, which we will trim in a moment. If your server is not in the catalog, click Enter a URL manually . You get the same connection with two fields: The manual path: an MCP server URL, optional Connection headers for keyed servers, and Connect. MCP server URL is the endpoint, https://mcp.deepwiki.com/mcp in our case. HTTPS is required, and a URL that resolves to a private or internal address is refused outright. Connection headers optional is where a key goes if your server wants one, for example a Key of Authorization and a Value of Bearer YOUR TOKEN . Quickchat AI stores the header and attaches it to every request on the server side, so the token never appears in a prompt or a conversation. What if the server needs a sign-in? Many of the most useful servers are account-bound: scheduling, payments, CRM. For those, there is no key to paste at all. Connect to a server like Calendly and Quickchat AI detects the OAuth challenge and walks you through it: An OAuth server detected: no key to paste, just Connect with Calendly and an approval popup. Click Connect with Calendly , approve access in the popup, and the tools appear, the same as the keyless flow. Under the hood this is OAuth 2.1 with dynamic client registration and PKCE: Quickchat AI registers itself with the server, exchanges one-time codes, stores the tokens encrypted, and refreshes them for you. You can Reconnect or Disconnect from the action at any time, and if you close the window mid-authorization, the connection resumes where you left off. We built a complete agent on an OAuth server in a separate guide: an AI scheduling assistant on Calendly’s MCP server https://quickchat.ai/post/ai-scheduling-assistant-calendly that checks real availability and books real meetings. Choose which tools the Agent may call The Tools panel is where you grant least privilege, and it is the most important screen in this guide. Open the action you just created Edit Action on its card and find the Tools section. It lists every tool the server publishes, with a switch per tool and a counter at the top. Least privilege in one panel: ask question on, the other two off, and Default tool activation off so future tools wait for your review. DeepWiki publishes three: | Tool | What it does | Hoist needs it? | |---|---|---| ask question | Given a repo and a question, returns a source-grounded answer | Yes | read wiki structure | Given a repo, returns its documentation outline | No | read wiki contents | Given a repo, returns its documentation | No | Switch on ask question only. The other two are not harmful, but every enabled tool is one more thing the model can reach for, one more description in its context, and one more behavior you have to test. A tool that is off does not exist for the model. Then turn Default tool activation off. That toggle decides what happens to tools the server adds in the future : off means a new tool stays disabled until you review it. This matters more than it looks. A remote server is software someone else updates; the tool list you approved today is not a contract. The npm world learned this the hard way when a well-reviewed MCP package shipped clean releases for months and then quietly added email exfiltration the postmark-mcp incident; more on it in the safety section . You cannot audit a third party’s release process, but you can make sure their new capabilities do not flow into your Agent silently. This toggle is that guarantee. On a Tuned by Quickchat server the same panel starts in the recommended state instead: connect Calendly and five of its 36 tools arrive enabled, with the rest off and Default tool activation already off. Use recommended resets to that curated set after you experiment. For larger servers the panel grows a search box, and Enable all and Disable all do what they say. A Tuned by Quickchat server arrives pre-trimmed: 5 of Calendly’s 36 tools enabled, and Use recommended resets the selection after you experiment. Tell the Agent when to reach for the tool The main prompt is the one place you control when the tool fires and what it asks about. Here is the part that surprises even people who know MCP well: the action’s name and description fields are labels for you, and the model never reads them. What the model reads is the tool descriptions the server publishes, plus your Agent’s main prompt. We verified this from the running code and with a hostile experiment, in the tuning section below. So the steering happens in exactly one place you control. Under Identity , paste this into the AI Main Prompt and swap in your own product details: You are the support assistant for Hoist, a resumable file upload API for developers. Hoist gives developers a hosted, resumable upload endpoint plus storage. Customers upload files straight from their own web and mobile apps using the open-source tus-js-client library, which speaks the tus resumable upload protocol to Hoist. Answer questions about Hoist itself plans and limits, API keys, storage regions, webhooks, retention from your knowledge. Be concise and friendly. Many Hoist customers ask how the tus-js-client library they install in their app actually behaves: how it resumes after a dropped connection, how it fingerprints a file, how to set the chunk size or retry delays, and which options it supports. That behavior lives in the library's source code, not in Hoist's docs. When a visitor asks a detailed question about how tus-js-client works, call the ask question tool with repoName "tus/tus-js-client" and their question, then answer using what it returns, relating it back to using Hoist. Do not use the tool for questions about Hoist's own product, pricing, or account settings. Two things in that prompt do the real work. It names the exact repository tus/tus-js-client , because repoName is an argument the model fills in on every call, and one sentence turns that inference into a constant the tuning section measures how well the inference holds up without it . And it scopes the tool to library-internals questions, so the Agent keeps answering Hoist’s own questions from its knowledge. One more thing the model sees, which nobody’s prompt shows: when an Agent has connected tools, Quickchat AI appends a short, standing Connected Action Tools section to the system prompt with general rules for using tools well. Your prompt sits on top of that policy, so you write the when and the what, and the platform covers the how. Test it, and watch the tool call happen Ask a question only the library’s source can answer, then open the conversation and read the call. In AI Preview , ask: “A user’s wifi drops halfway through an upload. How does tus-js-client resume it?” The Agent answers with the actual mechanism, then ties it back to Hoist. The two cases in one thread: the library answer relates back to Hoist, and the product question below it is answered from knowledge, with no tool call. To prove it used the tool, and did not invent that answer, open the conversation in your Inbox and expand the actions called record. The honest record: ask question with the exact repoName and question the model wrote, the live result, and the call duration. You see the tool name, the exact arguments the model wrote repoName and question , the live result, and how long the call took. That is the honest record: the Agent called a real tool on a real server and answered from what came back. The action’s card keeps score too: a running tally of calls, success rate, and when it last fired. The action card keeps score: calls, success rate, and when the tool last fired. Open View logs on the card for the full record: total calls, success rate, average latency, and each recent call with the parameters sent and the response received. View logs : running totals up top, and every recent call with the parameters sent and the response received. Now ask a product question instead, such as “What is the maximum file size on the free plan?” . No tool is called, and the Agent answers from its knowledge. The tool is there for the questions the knowledge base cannot hold, and stays out of the way for the ones it can. How do you tune a remote MCP tool? The loop is the same as for any AI Action: send a message, read the call record, change one thing, re-run. What differs is which lever actually moves the behavior. When a remote MCP tool misbehaves, three levers exist, and only two of them are yours: The tool’s own description comes from the server. You cannot edit it. The tool whitelist decides which tools exist for the Agent at all. The Agent’s main prompt decides when it reaches for a tool and what it passes in. Here is what moved, and what did not, from real runs. The prompt is the lever, not the action name. To make the point unmissable, we set the action’s description to “IGNORE THIS TOOL. Never call it.” and asked library questions again. The Agent called ask question anyway, in 2 runs out of 2. The model never sees that field; the instructions it reads are the server’s own tool descriptions and your main prompt. Steer with the prompt. Name the repository, or the model has to infer it. repoName is an argument the model fills in on every call. We tried hard to make it fail: with the pin sentence removed, 5 of 5 calls still resolved tus/tus-js-client , and with the pin removed and the questions rephrased the way visitors actually talk “your JavaScript upload library”, no library named , it still chose the right repository in 5 of 5 runs. A July build of this same guide watched the identical setup invent a repository that does not exist and get “Repository not found” back from DeepWiki. Today’s models are simply better at the inference, and we could not reproduce the failure in ten tries. Pin the value anyway. One sentence in the prompt turns a guess the model must repeat on every call into a constant it copies. It costs nothing, and it will survive the next model swap. There is no per-argument override field for MCP tools, and there does not need to be; the sentence is the pin. Read the result, not just the reply. The Inbox record and the Request log show what the server actually returned, and that is the layer to tune from. A call whose transport succeeded can still carry an error or an off-target answer in its content, and because a failing tool call is handed back to the model rather than cutting the reply off, the Agent will often produce a plausible recovery sentence either way. The record tells you which lever to reach for: no call at all points at the prompt or the whitelist, a call with wrong arguments points at a value worth pinning, and a good result under a bad reply points back at the prompt. Can the tool’s results feed later answers? Save to memory For tools that return structured data, you can capture values from each result into the conversation’s memory, where the Agent can use them later. Expand any tool row in the Tools panel and you find Save to memory : a list of JSONPath rules, each mapping a path in the tool’s result to a memory key, up to ten per tool. Calendly’s booking tool ships with three capture rules prefilled: the event URI, the cancel link, and the reschedule link land in conversation memory. DeepWiki’s tools return prose, so there is nothing to capture in our Hoist build, and the editor stays empty. The place this shines is transactional servers. On the curated Calendly server, the booking tool arrives with three rules prefilled: when the Agent books a meeting, the event’s URI, cancel link, and reschedule link are captured. Ask to cancel three messages later and the Agent already holds the exact cancel URL, no re-lookup, no guessing. Captured values become part of the conversation data the model reads, which is the closest thing MCP tools have to feeding one tool’s output into the next tool’s arguments. Two details for the security-minded, because this feature was built with a specific attack in mind. Captured values live in a namespace the server side owns, so a visitor cannot forge them by typing something that looks like a memory entry. And a capture that fails, because the result shape changed or the path matched nothing, never costs the model its tool result; the reply proceeds and the capture is simply skipped. Test it at scale with Simulations One conversation proves the tool fires. To check the Agent handles a whole set of questions the way you want, use Simulation Testing, under the Testing tab. Build a small dataset that mixes the two cases: library-internals questions that should reach for the tool, and product questions the Agent should answer from its knowledge. Give the evaluator your grading criteria, and include your real product facts in them, so it can tell a correct plan or region from an invented one. Then run the set and read the scores. The five-question run: 4.8 out of 5 overall, with a per-reply score and the judge’s justification for each. Our five-question run scored 4.8 out of 5: full marks on both product answers and on the resume and chunk-size explanations, with the retry answer docked to a 4. The evaluation grades the Agent’s reply, not the tool call itself, so pair it with the Inbox: the Inbox record proves the tool fired with the right inputs, and the Simulation shows the reply held up across many questions at once. Two costs to know before you run one. Every simulated message counts toward your plan’s monthly AI messages, the same as a real conversation. And a Simulation exercises your Agent for real, which means active MCP tools fire for real during a run . Harmless for a read-only tool like ask question ; something to think about before you simulate an agent whose tools write to a calendar or a payment system. Is connecting to an MCP server safe? Safety rests on the whitelist and on choosing the right kind of server, not on the prompt. Whitelist to least privilege. The Agent can call only the tools you switch on. Enable the one you need, and with Default tool activation off, a tool the server adds later stays off until you review it. For a read-only server like DeepWiki, the most a whitelisted ask question can do is read public data. Treat the server as a supply chain. You are trusting someone else’s code and someone else’s release process. Prefer first-party servers run by the company whose data they expose, read the tool list before enabling, and let the default-off toggle absorb whatever they ship next. The postmark-mcp incident, fifteen clean releases and then a BCC to an attacker in the sixteenth, is the case study to remember. Keep secrets out of the prompt. If your server needs a token, put it in a connection header, or use the OAuth flow where Quickchat AI holds encrypted tokens for you. Either way the credential is attached server-side, and the model never sees it, so no conversation can leak it. Internal addresses are blocked. Quickchat AI refuses to connect an MCP action to a private, loopback, or link-local address, and applies the same check to every endpoint an OAuth server advertises, so a server URL cannot be turned into a way to reach something inside a network. Slow servers cannot stall your Agent. Every connection is bounded: 10 seconds to connect, 20 seconds to read, inside a 20 second per-server budget per reply. A tool call that fails or times out is handed back to the model as an error, so the Agent explains what it could not check instead of dying mid-reply. Know the one honest limit. Unlike an API Action, an MCP tool has no per-conversation run condition: you cannot say “only run this tool when the visitor is a verified admin.” The whitelist is per tool, for everyone. If a server exposes a write or delete tool that you need to gate by who is asking, do not enable it here. Build that one operation as an API Action with a Run only when condition https://quickchat.ai/post/connect-ai-agent-to-any-api instead, and let the MCP connection carry the read-only work. And resist the urge to connect everything. Every enabled tool adds its description to what the model weighs on every reply, and published measurements show tool selection collapsing as the count grows: near-perfect at twenty tools, unusable past a hundred. The practitioner consensus of a handful of servers per agent matches what we see. One well-chosen server with one well-chosen tool, which is literally what this guide builds, outperforms a wall of switches. Have ChatGPT or Claude build this for you Everything above is clicking and pasting. If you would rather describe it than do it, hand this guide to an AI assistant that is connected to your Quickchat AI account. Quickchat AI has its own MCP server for managing your account: click Manage your AI right under the Agent switcher in the dashboard sidebar for connection tabs for ChatGPT, Claude, Claude Code, Cursor, and Codex. On ChatGPT you can also skip the dashboard entirely: install the Quickchat AI app for ChatGPT https://chatgpt.com/plugins/plugin asdk app 6a4656c688748191be4c5247fb0d5dfc directly and sign in when it asks. Once connected, the assistant can create agents, write knowledge bases, configure actions including MCP connections, read conversations, and run Simulations, with your role and confirmation gates enforced on every call. Manage your AI : connection tabs for ChatGPT, Claude, Claude Code, Cursor, and Codex. One sign-in, no API key. Then paste a prompt like this into the connected assistant: Using my Quickchat AI account, build the agent from the guide at quickchat.ai/post/connect-ai-agent-to-mcp-server. Create an AI Agent named Hoist for a resumable file-upload API for developers. Write it a short knowledge base: two plans Free: 5 GB storage, 2 GB max file, 30 day retention; Pro at $29/month: 500 GB storage, 50 GB max file, unlimited retention , API keys under Settings, EU and US storage regions, and an upload.completed webhook. Connect the remote MCP server https://mcp.deepwiki.com/mcp with only the ask question tool enabled. Set the main prompt so the agent answers Hoist questions from its knowledge, and calls ask question with repoName "tus/tus-js-client" for questions about how the tus-js-client library behaves. Then create a simulation dataset with two library questions and two product questions, run it, and report the scores. The MCP half of that flow runs on two tools Quickchat AI’s server exposes for exactly this: list remote mcp server tools previews the tools a server offers before anything is created, and create remote mcp action connects the server with a least-privilege allow-list, activated and ready. The assistant reports what it enabled and why, the same information you would read off the Tools panel. And when you then log in, Quickchat AI notices the account is already being managed over MCP and skips the Getting Started onboarding entirely — the dashboard opens straight on the agent your assistant built. One troubleshooting note for the ChatGPT path: if the sign-in popup opens blank, pause your ad blocker for chatgpt.com. A popular blocker heuristic breaks all ChatGPT connector popups, ours included, and the block happens entirely on ChatGPT’s page. If you would rather script it, the REST route needs exactly two calls with an API token https://docs.quickchat.ai/api-reference/ Business plan : POST https://app.quickchat.ai/v1/api/ai actions/remote mcp Authorization: Bearer YOUR API TOKEN {"name": "DeepWiki", "remote mcp url": "https://mcp.deepwiki.com/mcp", "headers": , "allowed tools": } PATCH https://app.quickchat.ai/v1/api/ai actions/ACTION ID Authorization: Bearer YOUR API TOKEN {"is active": true} The create returns the action with default tool is active on, so tools discovered at reply time work immediately; tighten the whitelist in the dashboard afterward, or pass allowed tools explicitly. Going live When the Agent behaves in AI Preview and the Inbox shows the calls you expect, deploy it. The MCP action belongs to the Agent, not to a channel, so the same tool works on your website widget, WhatsApp, Discord, Slack, and everywhere else you put the Agent. Keep in mind that every conversation, including the ones you run in AI Preview while testing, counts toward your plan’s monthly AI messages. One protocol, three directions “MCP” shows up in three different places in Quickchat AI, and mixing them up is the most common source of confusion we see in support conversations. This guide covered the first row: One protocol, three directions. This guide is the top lane; the other two rows have guides of their own. Connecting your Agent to a server gives it live tools, which is what we just built. The reverse direction, exposing your Agent as an MCP server https://quickchat.ai/post/expose-ai-agent-as-mcp-server , lets ChatGPT, Claude, and Cursor call your Agent as a tool and answer from your knowledge base. And managing your account over MCP https://quickchat.ai/post/manage-ai-agent-from-chatgpt is the build-it-with-AI route from the previous section, pointed at your dashboard instead of your visitors. Related guides The same AI Actions mechanism connects an Agent to any tool or API. Other step-by-step walkthroughs: Make your AI Agent’s actions reliable https://quickchat.ai/post/reliable-ai-agent-actions , the tuning discipline in depth Connect your AI Agent to Google Sheets https://quickchat.ai/post/connect-ai-agent-to-google-sheets Connect your AI Agent to HubSpot https://quickchat.ai/post/connect-ai-agent-to-hubspot For more on the protocol itself: GPT Actions versus MCP https://quickchat.ai/post/gpt-actions-vs-mcp , MCP versus plain HTTP APIs https://quickchat.ai/post/mcp-vs-http , and APIs for AI agents https://quickchat.ai/post/apis-for-ai-agents-from-mcp-to-custom-endpoints . Frequently asked questions What is a remote MCP server, and how is it different from a local one? A remote MCP server is a tool server you reach over HTTPS at a public URL, run by someone else, with nothing to install; consumer AI apps like ChatGPT and Claude surface the same servers as connectors in their settings. A local MCP server is a process running on your own machine, which only clients on that machine can use. A hosted, customer-facing AI agent needs the remote kind: a fixed URL it can call from the cloud on every conversation. Does MCP use HTTP? What happened to SSE? Yes. Since the 2026-07-28 revision of the spec, a remote MCP server is a plain, stateless HTTPS endpoint using streamable HTTP, and the older HTTP-plus-SSE transport is deprecated with a migration window. In practice you paste one URL. Quickchat AI speaks streamable HTTP and still falls back to SSE for legacy servers whose URL path ends in /sse. How does my AI agent find out what tools an MCP server offers? It asks the server. MCP defines a tools/list request: the client connects to the server URL and the server replies with each tool’s name, description, and input schema. In Quickchat AI this discovery runs when you connect, so the tools appear as a list of switches, and the agent re-reads the live tool list at reply time. Can I connect an MCP server without writing or hosting any code? Yes. Connecting is a catalog click, or pasting a URL, plus switching on the tools you want. Everything in this guide is a copy-paste value, and there is no code to write, host, or deploy on any plan. Is it free? Yes. MCP actions are available on every Quickchat AI plan, including the free one, and a connected MCP server counts as one action toward the limit of 15 AI Actions per agent. Creating API tokens for the REST route is the one part that needs the Business plan. Do I need OAuth, or are there MCP servers with no authentication? Both kinds exist. Documentation and search servers such as DeepWiki, Context7, and Microsoft Learn are open: paste the URL and you are connected. Account-bound servers such as Calendly, Notion, or Stripe use OAuth: instead of pasting a key, a popup opens, you approve access, and Quickchat AI stores the tokens. A plain API key in a connection header also works for servers that use one. I connected an MCP server. How do I actually use it? Tell the agent when to reach for it. Add one paragraph to the AI Main Prompt naming the situations where the tool helps and any values it should pass, then ask a matching question in AI Preview. The connection gives the agent the ability; the prompt gives it the occasion. Why is my agent not calling the MCP tool? Almost always for one of three reasons: the tool is switched off in the action’s Tools panel, the main prompt never tells the agent when to use it, or the question you asked is one the knowledge base already answers, so the agent has no reason to call out. Check the Tools panel first, then add explicit guidance to the prompt, then test with a question the knowledge base cannot possibly answer. How many MCP servers should I connect to one agent? Fewer than you think. Every enabled tool’s description is loaded for the model on every reply, and published tests show tool selection degrading as the count grows, with near-perfect accuracy at around 20 tools collapsing entirely past 100. Practitioners converge on a handful of servers per agent. Connect the one or two your agent genuinely needs and switch on only the tools you use. How do I limit what a connected MCP server can do? Switch on only the tools the agent needs in the Tools panel and turn Default tool activation off, so a tool the server adds later stays off until you review it. The agent can never call a tool you have not enabled. Keep credentials in connection headers rather than the prompt, and gate genuinely sensitive write operations behind an API Action with a run condition instead of an MCP tool. Can ChatGPT or Claude build this Quickchat AI setup for me? Yes. Connect your AI app to Quickchat AI’s own MCP server from the dashboard, then ask it to create the agent, write the knowledge base, connect the MCP server, and run the tests. There is also a REST route: creating an MCP action is one POST plus one PATCH with an API token. What happens if the MCP server is slow or down mid-conversation? The conversation does not hang. Quickchat AI bounds every server with short timeouts, 10 seconds to connect and 20 seconds to read, inside a 20 second per-server budget for each reply. A failed or timed-out call is handed back to the model as an error, so the agent answers without the tool and explains what it could not check. The Inbox records the failed call.