ChatGPT plugin for AutoSEO: how to connect it via MCP

What happened to ChatGPT plugins?

The original plugins, launched in March 2023, were APIs described by an OpenAPI spec and a manifest file (OpenAI, 2023). OpenAI retired them in spring 2024 in favor of GPTs (OpenAI Help Center, archived 2024).

What followed runs on MCP. In September 2025 OpenAI announced developer mode, full MCP client support for read and write tools (OpenAI Developer Community, 2025), then apps and the Apps SDK (OpenAI, 2025). In July 2026 the App Directory became the Plugin Directory (ChatGPT release notes, 2026). A plugin today can bundle skills, an MCP server and optional UI, and ChatGPT and Codex share one directory (OpenAI, 2026). For AutoSEO, the part that matters is the MCP server.

What do you need before you connect AutoSEO?

• The MCP URL. In AutoSEO, Settings → API & MCP shows your instance URL plus /api/mcp in the MCP Server panel. On AutoSEO Cloud you copy it from your workspace; a self-hosted instance uses your own domain.
• Public HTTPS for ChatGPT. ChatGPT only connects to remote MCP servers (OpenAI Help Center, 2026), and AutoSEO's settings page warns when an instance runs on localhost. Codex, Cursor and VS Code run on your machine and can reach a local instance.
• An AutoSEO role that may connect apps. Approving an OAuth connection requires the workspace permission Integrations, API keys, model settings, the same one needed for API keys.
• A ChatGPT plan with developer mode. OpenAI's developer docs list Plus, Pro, Business, Enterprise and Edu on the web (OpenAI, 2026). The Help Center says full MCP support including writes is in beta for Business, Enterprise and Edu, that Pro users can connect servers with read/fetch permissions, and that workspace admins enable developer mode first. The pages don't fully agree, so check what your account shows.

How do you add AutoSEO to ChatGPT?

As of September 2026, OpenAI documents this flow (OpenAI, 2026):

1. On chatgpt.com, open Settings → Security and login and turn on Developer mode. In Business, Enterprise and Edu workspaces an admin may have to allow it first under Workspace settings → Permissions & roles.
2. Go to ChatGPT Plugins and select the plus button to create a developer-mode app for a remote MCP server.
3. Name it “AutoSEO”, add a short description and paste your MCP URL under Connection.
4. Choose OAuth as the authentication. ChatGPT reads AutoSEO's metadata, registers itself and sends you to AutoSEO's consent page.
5. Pick the workspace, the projects (or All my projects) and the permissions, then click Allow access. ChatGPT then scans the tools.
6. In a new chat, open the + menu, choose Developer mode and select AutoSEO, or @-mention it.

Older guides, including the steps in AutoSEO's own settings page, still say Settings → Apps & Connectors. OpenAI's current docs place the toggle under Security and login, and apps became plugins in July 2026, so follow OpenAI's wording if menus don't match.

How do you connect Codex, Cursor and VS Code?

These clients run locally, so an API key is the simplest route. Create one under Settings → API & MCP → API Keys (keys start with as_live_), name it after the client and keep it in an environment variable, not in a file you might commit.

OpenAI Codex (CLI, IDE extension, desktop app)

Codex stores MCP servers in ~/.codex/config.toml, shared by the CLI, the IDE extension and the ChatGPT desktop app. An HTTP server takes a url and an optional bearer_token_env_var; OAuth runs through codex mcp login (OpenAI Codex docs, 2026).

codex mcp add autoseo --url https://your-autoseo.example.com/api/mcpcodex mcp login autoseo

# export AUTOSEO_API_KEY="as_live_..." in your shell profile first[mcp_servers.autoseo]url = "https://your-autoseo.example.com/api/mcp"bearer_token_env_var = "AUTOSEO_API_KEY"default_tools_approval_mode = "writes"

In writes mode Codex asks before any tool not marked read-only. AutoSEO marks every tool that changes data or can cost money that way, so reads run freely and changes wait for you.

AutoSEO also serves a plugin bundle with the MCP server and 17 skills, such as a site audit and an AI visibility report. Download <your instance URL>/api/plugin/autoseo-marketplace.zip, unzip it, run codex plugin marketplace add ./autoseo-marketplace and install AutoSEO from /plugins in the Codex CLI; plugins aren't available in the IDE extension (OpenAI, 2026). See AI SEO agent for what the skills do.

Cursor

Add the server to ~/.cursor/mcp.json (all projects) or .cursor/mcp.json (one project). Cursor interpolates environment variables, supports OAuth for remote servers, asks before running MCP tools by default and toggles servers in the Customize sidebar (Cursor docs, 2026).

{  "mcpServers": {    "autoseo": {      "url": "https://your-autoseo.example.com/api/mcp",      "headers": { "Authorization": "Bearer ${env:AUTOSEO_API_KEY}" }    }  }}

VS Code

For GitHub Copilot's agent mode, add .vscode/mcp.json or run MCP: Add Server. An input with password: true makes VS Code prompt for the key instead of storing it in the file (VS Code docs, 2026).

{  "servers": {    "autoseo": {      "type": "http",      "url": "https://your-autoseo.example.com/api/mcp",      "headers": { "Authorization": "Bearer ${input:autoseo-api-key}" }    }  },  "inputs": [    { "type": "promptString", "id": "autoseo-api-key", "description": "AutoSEO API key", "password": true }  ]}

What can you ask once AutoSEO is connected?

At the time of writing, AutoSEO's MCP server registers more than 110 tools, grouped by area from AI visibility and competitors to Search Console, site audits, reports and CMS edit proposals; Settings → API & MCP lists each with its scope. The server tells the model to start with list_projects and get_project_context, so you rarely need a project ID.

Visibility numbers come from AutoSEO's own prompt tracking (ChatGPT visibility tracking); scripts and dashboards get the same data from the REST API.

How do you keep the connection safe?

AutoSEO hides what a credential can't use: tools/list returns only tools whose scope the connection holds, and every call is checked again against scope and role.

• Read: projects, prompts, visibility metrics, competitors, sources, tasks and reports; always included. A read-only credential sees just over half of the tools, none of which change data or spend money.
• Write: create projects, add prompts, update tasks, save reports and propose CMS edits, bounded by your role. CMS proposals change nothing on the live site until a project member applies them.
• Spend credits: every tool that can cost money, such as DataForSEO research, AI generation, audits and tracking runs. They also need the matching role permission.
• Export: bulk export of prompts, answers and results through the REST API; no MCP tool requires it today.

For analysis in ChatGPT, approve only Read on the consent screen and limit the grant to the projects you need. Add Write when the assistant should change things, and Spend only for sessions with planned paid research.

Every AutoSEO tool that writes or can incur cost is published without readOnlyHint, so ChatGPT asks before running it by default. Tool results contain third-party text such as AI answers, so treat them as data; OpenAI warns about prompt injection through connected servers (OpenAI Help Center, 2026). Whether OpenAI may train on what ChatGPT reads from apps depends on your plan and settings (OpenAI Help Center, 2026).

To revoke access, open Settings → API & MCP: Connected apps → Disconnect immediately revokes all tokens of that app, and Revoke key ends an API key. OAuth access tokens last one hour, refresh tokens roll over 30 days, and reusing a rotated refresh token revokes the whole grant.

What if the connection doesn't work?

What does a first session look like?

Example (hypothetical): a marketing lead at Northwind Tools, a fictional B2B software company, uses ChatGPT Plus and a self-hosted AutoSEO that tracks 60 prompts across ChatGPT, Perplexity and Google AI Overviews. All numbers are illustrative, not real data.

1. She adds AutoSEO in developer mode and grants Read for one project only.
2. “Compare our ChatGPT visibility for the last 30 days with the 30 before.” ChatGPT calls list_projects, then get_visibility_metrics: visibility fell from 24% to 19% (illustrative).
3. “Where are competitors named instead of us?” get_competitor_gap_analysis returns 11 prompts (illustrative) naming one competitor but not Northwind, plus two review sites that cite only that competitor.
4. “What did ChatGPT search for on those prompts?” get_query_fanouts shows sub-queries about pricing and alternatives, pointing to a missing comparison page.
5. list_tasks surfaces a matching content task. In Codex, where her key has Read and Write, she asks to set it to in progress; Codex pauses for approval because update_task_status is a write tool.

The session never needed Spend: all data came from tracking runs AutoSEO had already completed.

What does this mean for you?

• Treat anything mentioning an OpenAPI manifest or the plugin store as the retired 2023 system; current ChatGPT plugins are MCP-based.
• Connect ChatGPT with OAuth and a read-only, project-limited grant first; widen it only when a task needs it.
• Give each local client its own named API key in an environment variable, and revoke keys you no longer use.
• Leave Spend credits off by default and keep write confirmations on.
• Recheck the setup when ChatGPT's menus change: in three years the names went from plugins to GPTs, connectors, apps and back to plugins.
• Keep two questions apart: this connection lets ChatGPT read your AutoSEO data, while ChatGPT visibility tracking measures how ChatGPT talks about your brand. The Claude connector guide covers the same setup for Claude.