Free tools Windows power users keep installed
One-click scans. No signup required.
Built-in MCP servers are first-party integrations that let an OpenAI client discover and call tools, read resources, and use prompts supplied by a Model Context Protocol (MCP) server. “Built-in” describes support in the host product—not a promise that every third-party server is bundled or available on every plan. Codex supports local STDIO processes and remote streamable HTTP servers; ChatGPT developer mode connects to remote servers; and the Responses API can call public remote servers or private ones through Secure MCP Tunnel.
OpenAI also publishes a read-only documentation server at https://developers.openai.com/mcp. The sections below show where each option fits, how to add the documentation server to Codex, what changes in ChatGPT and the API, and how connectors, plugins, and MCP servers differ.
What an MCP server provides
MCP is an integration layer between an AI client and an external service. An MCP server can publish four kinds of capabilities:
- Tools that the model can call. Each tool has a name, description, input schema, and optional output schema.
- Resources that provide retrievable context.
- Prompts that package reusable instructions.
- Server instructions that describe how the service should be used.
The formal OpenAI description is in the MCP server concept documentation. A server remains an external process or service; the model does not host its code. The client discovers the server’s advertised capabilities and decides, subject to permissions and approvals, when to invoke them.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallCrashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minute#1 Best Overall
“Built-in” versus bundled
OpenAI has first-party MCP support in Codex, ChatGPT developer mode, plugins, and the API, and it hosts an official documentation server. That support is not the same as bundling every community or commercial server. You still select a server, configure its transport, provide credentials when required, and grant the permissions appropriate to your task.
Which OpenAI host supports which MCP setup?
| Host | Transport and reachability | Permissions and authentication | Important qualification |
|---|---|---|---|
| Codex CLI, desktop, and IDE extension | Local STDIO processes or remote streamable HTTP URLs | Streamable HTTP can use bearer tokens or OAuth, including Client ID Metadata Documents and Dynamic Client Registration. | The ChatGPT desktop app, Codex CLI, and IDE extension share MCP configuration for the same Codex host, according to ChatGPT Learn. |
| ChatGPT developer mode | Remote MCP servers. A local or private server needs Secure MCP Tunnel. | Plan and app permissions determine whether read/fetch or write/modify actions are available. | Full write/modify support is rolling out in beta to ChatGPT Business, Enterprise, and Edu. Pro users can connect MCPs with read/fetch permissions. OpenAI-built apps are search-only today, as documented by the Help Center. |
| Responses API | Any public server on the Internet that implements remote MCP; Secure MCP Tunnel can reach a local or private server. | Use the server’s bearer/OAuth flow and require approval when data should not be shared automatically. | The API guide at Tools, connectors, and MCP says the legacy connector_id path is deprecated for models released after September 1, 2026; existing models retain connector support. |
| Plugins | Plugin-hosted MCP server capabilities. | OAuth 2.1 is recommended when a plugin needs user authentication. | A plugin can expose tools without a user interface or add an optional web component rendered in ChatGPT. See the plugin app quickstart. |
The official OpenAI Docs MCP server
OpenAI hosts a public, documentation-only MCP server at https://developers.openai.com/mcp. It offers read-only search and page-content access for developers.openai.com, platform.openai.com, and learn.chatgpt.com. It cannot modify those sites.
For Codex, the documented command is:
codex mcp add openaiDeveloperDocs --url https://developers.openai.com/mcpRun it in the environment where Codex is installed. The server name, openaiDeveloperDocs, is a local label; you may choose another name, but keeping the documented label makes examples and team instructions easier to compare.
Verify the connection
- Start a new Codex session after adding the server, or restart the current session so it reloads MCP configuration.
- Ask Codex to search the OpenAI documentation for a narrowly scoped question, such as a current Responses API parameter.
- Check that the answer uses pages from the three documented domains and that no write operation is offered. The server is read-only by design.
The OpenAI Docs MCP guide also shows equivalent entries for ~/.codex/config.toml and VS Code. Use those examples when you need to provision the same server through a checked-in configuration or an IDE rather than the CLI command.
How to add another MCP server to Codex
Choose the server’s transport first. A local server is a process that Codex starts over STDIO; a hosted server exposes a streamable HTTP URL. Then obtain the credentials and permission policy specified by that server’s documentation.
- For a local STDIO server: install its runtime and dependencies, and identify the executable plus arguments that start the server. Keep secrets in the process environment or the server’s secret store, not in prompts.
- For a streamable HTTP server: copy its HTTPS endpoint and determine whether it requires a bearer token or OAuth. Confirm whether it supports Client ID Metadata Documents or Dynamic Client Registration before choosing an OAuth flow.
- Register it with the Codex host: use the server’s Codex configuration instructions, selecting the local STDIO or remote HTTP transport. The OpenAI Docs example above is the complete remote command; other servers may require additional authentication fields.
- Restart and test: have Codex list or describe the available tools, then make a harmless read-only call before enabling any write capability.
Because the desktop app, CLI, and IDE extension use the same Codex host configuration, adding a server once can make it available in each of those clients. Availability still depends on the client version, the server’s authentication, and the permissions you grant.
Rank #2
Can ChatGPT use a local MCP server?
Not by connecting directly to a process on your laptop. ChatGPT developer mode connects to remote MCP servers. To keep a server private or local, put it behind Secure MCP Tunnel, which provides the route without requiring a public deployment.
Developer mode permissions
- Organizations can build, test, and deploy MCP-powered apps in developer mode.
- Full write/modify support is rolling out in beta for ChatGPT Business, Enterprise, and Edu.
- Pro users can connect MCPs with read/fetch permissions in developer mode.
- OpenAI-built apps are search-only today and do not support write actions.
These are product and plan capabilities, not properties of the MCP protocol itself. A server may advertise write tools, yet ChatGPT can still withhold them because of the account, app, or approval policy.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Using MCP from the Responses API
The Responses API accepts a remote MCP tool configuration for any server on the public Internet that implements remote MCP. For a server running on a private network or your workstation, use Secure MCP Tunnel instead of exposing an unauthenticated endpoint.
Authentication and approvals
Remote streamable HTTP servers commonly use bearer tokens or OAuth. Treat the server as an external service: review the inputs it receives, inspect its outputs, and require approval before sharing sensitive data or allowing a consequential action. The API guidance recommends establishing connections only with trusted servers and using approvals where appropriate; read the current API guide for the request shape supported by your model and SDK version.
Connector lifecycle
Older API integrations used a connector_id. For models released after September 1, 2026, that legacy path is deprecated; models released earlier retain connector support. Prefer the current remote MCP configuration for new work and check the model-specific API documentation before migrating an existing integration.
Connectors, plugins, and remote MCP servers: the difference
| Term | What it is | Where it runs | Typical permissions | What to watch |
|---|---|---|---|---|
| MCP server | A service or process that advertises tools, resources, prompts, and instructions. | Local STDIO, a streamable HTTP service, or a private endpoint reached through Secure MCP Tunnel. | Whatever the client and server jointly allow, from read-only retrieval to writes. | Validate schemas, outputs, credentials, and trust. |
| Remote MCP server | An MCP server reachable at a network URL. | Public Internet or private network through a tunnel. | Bearer/OAuth authentication and client approvals can limit calls. | Network exposure, token scope, TLS, and data residency. |
| Plugin | An application package that uses MCP to provide server-backed capabilities to ChatGPT or Codex. | Plugin infrastructure, optionally with a web component in ChatGPT. | Tool calls plus any UI actions the plugin implements. | Use OAuth 2.1 when users must authenticate; review both tool and UI behavior. |
| Connector | A client integration path, including the older API connector_id mechanism. |
Defined by the host or API rather than by a separate protocol. | Determined by the connector and host. | Legacy connector IDs are deprecated for models released after September 1, 2026. |
In short, MCP is the capability protocol, a remote MCP server is one deployment style, a plugin is an app wrapper, and a connector is a host/API integration mechanism. They can overlap, but the terms are not interchangeable.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Security checklist before enabling a server
- Trust the operator: a tool can receive the arguments and context you send it, so use servers from operators you can evaluate.
- Minimize scope: prefer read/fetch permissions until a write action is necessary, and grant the narrowest OAuth or bearer-token scope available.
- Use approvals for sensitive calls: require a human confirmation before sending private records, changing production data, publishing content, or making purchases.
- Inspect schemas and outputs: tool descriptions are metadata, not a security guarantee. Validate returned URLs, files, and structured values in your application.
- Keep local services private: use Secure MCP Tunnel for ChatGPT or the API rather than opening an unauthenticated port to the Internet.
- Rotate credentials: revoke tokens when a server, workstation, or team member no longer needs access.
For plugin-specific authentication, OpenAI recommends OAuth 2.1 when user authentication is required; the relevant guidance is in the plugin MCP server documentation.
Troubleshooting common MCP failures
Codex cannot find the server
Confirm that you ran the registration command in the same Codex host environment you are using, then restart the session. For a local server, run its startup command directly and check that it stays alive and writes protocol messages to the expected STDIO streams rather than mixing logs into the protocol.
A remote server returns an authentication error
Verify the HTTPS URL, token audience and scope, or OAuth redirect and registration details. Streamable HTTP supports bearer tokens and OAuth, but the server decides which flow and scopes are valid. Do not paste a secret into a prompt to “fix” authentication.
ChatGPT cannot connect to a localhost URL
ChatGPT developer mode requires a remote endpoint. Place the private service behind Secure MCP Tunnel, then register the tunnel address and its authentication instead of exposing your laptop directly.
A write tool is visible but cannot run
Check the account plan, developer-mode permission, app policy, and approval requirement. ChatGPT Pro supports read/fetch MCP permissions, while full write/modify support is rolling out in beta to Business, Enterprise, and Edu. OpenAI-built apps remain search-only today.
The documentation server does not answer a product question
It only searches and returns page content from developers.openai.com, platform.openai.com, and learn.chatgpt.com. It is not a general web search service and has no write tools. Ask for a page or API concept within those domains, or connect a different trusted server for other sources.
Rank #4
An old API integration stopped working after a model change
If it depends on connector_id, check the model release date. Connector IDs are deprecated for models released after September 1, 2026, although existing models retain support. Migrate to the current remote MCP configuration where the model documentation requires it.
A practical external MCP server: ScreenshotNeo
ScreenshotNeo is not bundled into OpenAI products, but it is a useful MCP server for agents that need website captures. Its MCP tools are take_screenshot, get_page_info, and capture_pdf; connect it as an external service according to its documentation.
ScreenshotNeo also exposes a direct HTTP API. One GET request returns a PNG, JPEG, WebP, or PDF. Before capture it accepts cookie and consent banners like a visitor and removes more than 60 known consent platforms, newsletter popups, and chat widgets; each cleanup step can be disabled. Bot checks or CAPTCHAs, blank pages, timeouts, failed loads, and cache hits are not billed, and response headers identify the page verdict and billing result.
Or skip the browser setup:
Use the one-call API when you do not need to manage a browser process yourself. The parameter names used by other screenshot APIs also work, which can simplify migration.
cURL — see the ScreenshotNeo API documentation for all options:
curl -G "https://api.screenshotneo.com/v1/shot" -d access_key=YOUR_API_KEY --data-urlencode url=https://stripe.com -o shot.webpPython:
import requests
r = requests.get("https://api.screenshotneo.com/v1/shot", params={"access_key": "YOUR_API_KEY", "url": "https://stripe.com"}, timeout=90)
open("shot.webp", "wb").write(r.content)Node.js:
const q = new URLSearchParams({ access_key: 'YOUR_API_KEY', url: 'https://stripe.com' });
const res = await fetch(`https://api.screenshotneo.com/v1/shot?${q}`);Every plan includes full-page captures with lazy images loaded, element capture by CSS selector, dark mode, device presets or custom viewports, retina scale, PDF controls, custom CSS and JavaScript, click-before-capture, selector hiding, waits, request blocking, headers, cookies, user agents, Authorization, timezone, geolocation, transparent backgrounds, resizing, configurable caching, signed links, asynchronous jobs with signed webhooks, bulk capture for up to 100 URLs per call, a usage API, and an OpenAPI specification. The free plan includes 1,000 screenshots per month without a card; paid plans start at $5 for 3,000. Clean shots are the only shots billed. An MCP server lets Claude, Cursor, or another MCP client call the capture tools.
Create a free ScreenshotNeo account to get 1,000 screenshots a month with no card.
Performance, reliability, and operating cost
Choose transport for the workload
STDIO avoids network exposure and is convenient for a process on the same machine, but the process must be installed and kept healthy. Streamable HTTP is easier to share across machines and clients, while adding network latency, authentication, TLS, and service-availability dependencies. A tunnel preserves private deployment at the cost of another managed connection.
Make calls predictable
- Use narrow tool descriptions and schemas so the model can select the right operation.
- Separate read tools from write tools and require approval for the latter.
- Set request, tool, and overall task timeouts in the client or application.
- Log tool name, sanitized arguments, result status, and latency without recording secrets.
- Cache stable resources where the server permits it, but do not cache user-specific or rapidly changing data without an explicit freshness policy.
Budget for the whole path
MCP itself has no single universal price. Your cost can include the host model, the server’s subscription or usage fees, network transfer, tunnel infrastructure, and any downstream API calls made by a tool. For the OpenAI Docs server, the documented capability is read-only documentation search and page retrieval; for commercial servers, read their plan and billing terms before enabling bulk or write operations.
FAQ
Frequently Asked Questions
Is an MCP server the same thing as an AI model?
No. The model generates decisions and arguments; the MCP server executes external tools or returns resources according to its own implementation.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsCan one MCP server be used by more than one AI client?
Yes, when each client supports the server’s transport and authentication. Codex clients sharing the same host configuration can use the same registered server, while ChatGPT and the API require a reachable remote endpoint or Secure MCP Tunnel.
Who controls whether a tool can make changes?
The server advertises its tools, but the client, account plan, app policy, authentication scope, and approval settings can restrict or block write actions.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
