ChatGPT Guide (2026): MCP, Apps & Automation

ChatGPT is the most widely used AI assistant on the planet in 2026: 900 million weekly active users, 50 million paying subscribers, and a $40 billion annualized revenue run rate as of August.¹² For marketers and agencies the more interesting shift is structural. ChatGPT is no longer just a chat box. It has a plugin directory, an app layer built on the Model Context Protocol, a scheduled task runner, and a long-running agent called ChatGPT Work. Sam Altman framed the change at DevDay 2025 as moving from "systems that you can ask anything to, to systems that you can ask to do anything for you."³ For social media teams, "do anything" now includes writing, validating, and scheduling posts through tools like posterly.

This is a long read (about 16 minutes). The table of contents on the right lets you jump. Every stat has a footnote so you can verify it, and we flag which features were retired this year (agent mode, the App directory name, and the Atlas browser) so you are not following stale advice.

What is ChatGPT?

ChatGPT is OpenAI's consumer and business AI assistant. In 2026 it runs across four surfaces: the web app, iOS and Android apps, and a desktop app for macOS and Windows that now contains "Chat and Work, plus Codex". Codex, OpenAI's coding agent, "is not selectable on web or mobile", so the desktop app is where the full set lives.⁴ A fifth surface, the ChatGPT Atlas browser, launched on October 21, 2025 and was deprecated on July 9, 2026; OpenAI said "Atlas is scheduled to stop working on August 9, 2026" and pointed users to the desktop app and the Chrome extension and sidebar instead.⁵

Models

The current flagship family is GPT-5.6, split into three named models: Luna, Terra, and Sol. They launched publicly on July 9, 2026 after a US Commerce Department review, and OpenAI's release notes say "GPT-5.6 Sol is rolling out to eligible paid ChatGPT plans. Free, Go, and logged-out users are not included."⁶⁷ Earlier in the year OpenAI shipped GPT-5.3 Instant (March 3), GPT-5.4 Thinking (March 5), a GPT-5.4 mini fallback (March 18), and a GPT-5.5 Instant update (May 28), while retiring GPT-5.1 on March 11 and GPT-4o, GPT-4.1, and o4-mini on February 13. The o3 model retires on August 26, 2026.⁶ Enterprise and Edu workspaces see a model picker with Instant, Medium, High, Extra High, Pro Standard, and Pro Extended effort levels and "virtually unlimited messages for eligible base models".⁸

Pricing

Two things trip up buyers. First, the OpenAI API is billed separately from every ChatGPT plan, so a Business seat does not include API credits.⁹ Second, Codex-only seats were closed to new workspaces after June 24, 2026, and Premium seats had a waitlist before August 20, 2026.⁹ Codex itself crossed 3 million weekly users on April 8, 2026 and spans a CLI, an IDE extension, a cloud runner, and code review.¹⁰

Key ChatGPT statistics in 2026

CFO Sarah Friar summed up the enterprise swing on August 14: "We entered the year at 60-40, but enterprise has accelerated much faster than expected and those lines have now crossed."² Third-party aggregators also circulate figures for paying business users and business customers, but we could not verify those against an OpenAI source, so they are deliberately left out of this table.

How ChatGPT connects to tools: MCP, apps and plugins

The Model Context Protocol is the open standard that lets an AI client discover and call tools on a server, whether that server runs on your laptop (stdio) or on the web (Streamable HTTP). ChatGPT's tool layer is built on top of it: OpenAI's Apps SDK, announced October 6, 2025 with launch partners including Booking.com, Canva, Coursera, Figma, Expedia, Spotify, and Zillow, "builds on the Model Context Protocol (MCP)".¹¹ OpenAI's stated intent was portability: "The Apps SDK is open source so that apps built with it can run anywhere that adopts this standard."¹²

The vocabulary changed twice in a year, which is why older tutorials feel confusing:

• Connectors became apps. Apps can add interactive UI, search, deep research, sync, and write actions inside a chat.¹³
• The App directory became the Plugin directory on July 9, 2026. OpenAI now says "A plugin can include skills, apps, and app templates". The directory is visible on all plans and usable in ChatGPT web, desktop, Work, and Codex.¹⁴ Developer docs for the Apps SDK now live under "Plugins", and submissions go through the OpenAI Platform.¹⁵

Custom MCP servers and developer mode

If you want to add a server that is not in the directory, you use developer mode. OpenAI's own words: "Take action with your company's tools and systems from ChatGPT, moving beyond read/search by creating apps with interactive UI and apps with full MCP support - including write/modify actions."¹⁶ The gating matters:

• Apps, full MCP support, and developer mode are available for ChatGPT Business and Enterprise/Edu customers on ChatGPT web.¹⁶
• Pro users can connect MCP servers with read/fetch permissions in developer mode. Write actions are Business and Enterprise/Edu only.¹⁶
• Web only. OpenAI's FAQ answers "Are MCP apps available on mobile? No."¹⁶
• Remote servers only. "Can I connect to a local MCP server? Not directly. ChatGPT connects to remote MCP servers. If your MCP server runs on a private network, on-premises, or on a developer machine, use Secure MCP Tunnel."¹⁶
• Servers no longer need to expose search and fetch tools to be accepted, and deep research "can use custom apps, but for read/fetch actions only".¹⁶

Codex is the exception

Codex, whether you use the CLI, the IDE extension, or the Codex tab in the ChatGPT desktop app, reads ~/.codex/config.toml (or a project-level .codex/config.toml) and does support stdio servers. OpenAI notes that "The ChatGPT desktop app, Codex CLI, and IDE extension share this configuration."¹⁷ You add servers with codex mcp add <name> --env VAR=VAL -- <command> for stdio, or codex mcp add <name> --url https://... for Streamable HTTP, and check them with codex mcp list.¹⁷

The API side

Developers building their own agents on OpenAI models get MCP in two places. The Responses API has an mcp tool that takes a server_url and a require_approval setting ("never", "always", or per tool), plus eight OpenAI-hosted connectors for Dropbox, Gmail, Google Calendar, Google Drive, Microsoft Teams, Outlook Calendar, Outlook Email, and SharePoint, with "no additional fees involved per tool call".¹⁸ The Agents SDK exposes a hosted MCP tool alongside MCPServerStreamableHttp and MCPServerStdio classes.¹⁹ Custom GPT Actions still exist too, defined with an OpenAPI schema and None, API key, or OAuth auth, but "A GPT can use either apps or actions, but not both at the same time" and "Actions are not available for Pro mode".²⁰

Connect ChatGPT to posterly

posterly ships one MCP server with two transports: a local stdio package on npm and a hosted HTTP endpoint. Because ChatGPT web only speaks to remote servers, the hosted endpoint is the one you want. Do not try to paste an npx config into ChatGPT web; there is nowhere for it to run. The full canonical walkthrough lives on the posterly MCP page and in the MCP docs.

Step 1: turn on API access in posterly

Sign up for any posterly plan, then open Dashboard, then API and MCP and enable the API and MCP add-on. It costs $3/month on top of any plan; the Agency API tier is $29/month. Click "Create Key" and copy the bearer key (it starts with pst_live_). Keys are scoped, and user-created keys are tier-based: Starter 1, Pro 2, Power 3, Agency 4. If you would rather let an agent do the signup for you, the agents signup page and the public start_signup tools handle it, which we cover in how AI agents sign users up to SaaS.

Step 2: add posterly as a custom MCP app in ChatGPT

You need a ChatGPT Business or Enterprise/Edu workspace for write actions, or Pro for read-only. Open ChatGPT on the web, enable developer mode, and create a new app pointing at the hosted endpoint:

Server URL:  https://www.poster.ly/api/mcp
Auth:        Authorization: Bearer pst_live_your_key_here
Wire format: JSON-RPC 2.0 over HTTPS (single request, no SSE in v1)

Two notes from posterly's docs. GET https://www.poster.ly/api/mcp returns server info without auth, which is a quick way to confirm the URL is reachable. And if your workspace only offers OAuth for custom apps, use the Smithery managed connection, which fronts the same hosted endpoint with posterly OAuth so you never paste a raw key. Either way the tool set is identical: 75 hosted HTTP MCP tools, listed on the MCP page and in the docs.

Step 3: verify with a read-only call

Ask ChatGPT: "Using posterly, run whoami and then list_accounts." whoami returns the user, workspaces, and scopes your key can access; list_accounts returns your connected social accounts. If you get a 401, the key is wrong; a 403 insufficient_subscription means the add-on is not active, and posterly's WWW-Authenticate header points ChatGPT at the exact upgrade URL.

Step 4: approve write actions

By default ChatGPT uses the "Important actions" permission level: reads run automatically, and it asks before consequential actions such as "Sending or editing an email, message, comment, post".¹³ Keep that on. When ChatGPT calls create_post you will be asked to confirm, which is exactly the safety net you want on a tool that publishes to a client's LinkedIn page. Note that after admin approval on Business plans, "ChatGPT uses a 'frozen' snapshot of its available tools", new tools are disabled by default, and published Business apps cannot be edited after publish (you recreate them), so if posterly adds a tool later an admin may need to re-approve.¹⁶

Codex users: local stdio works too

If you live in the ChatGPT desktop app's Codex tab or the Codex CLI, you can run the npm stdio server instead:

codex mcp add posterly --env POSTERLY_API_KEY=pst_live_your_key_here -- npx -y posterly-mcp-server@latest
codex mcp list

That writes an [mcp_servers.posterly] table into ~/.codex/config.toml.¹⁷ The stdio package exposes 78 tools, the 75 hosted tools plus the public setup and signup tools, and it can hand raw files to upload_media, which is the main reason to prefer it if you push a lot of video. Agents can also discover both transports automatically from https://www.poster.ly/.well-known/mcp/server-card.json.

What you can automate

Once connected, ChatGPT can drive posterly's tools with plain prompts. OpenAI's own plugin prompt library includes "Draft a social post for LinkedIn based on our brand tone and voice guide", and the OpenAI Academy marketing track suggests prompts like "Draft a 3-post social media series promoting [event, product, or milestone]" and "Create 5 ad copy variations for a [Facebook/LinkedIn/Instagram] campaign".²¹²² With posterly attached, those drafts become scheduled posts. Eight workflows we see agencies run:

1. Schedule a week of LinkedIn posts. Prompt: "Draft five LinkedIn posts for next week from this product update, run generate_captions for the brand voice, use find_available_slot for each, validate_post, then create_post after I approve." find_available_slot respects posterly's one-hour gap rule so nothing stacks.

2. Repurpose one blog into five platforms. Prompt: "Take this article and produce platform-native versions for X, Threads, Bluesky, Mastodon, and Facebook, then schedule them across Tuesday and Thursday." Use create_posts_batch to create up to 25 posts in one confirmed request instead of 25 separate calls.

3. Upload media and caption it. Prompt: "Pull the image at this public URL into posterly with upload_media_from_url, write an Instagram caption with three hashtags, and schedule it for 9am Friday." Hosted HTTP MCP is JSON only, so public URLs or the stdio route are the right path for larger files.

4. Pull last month's analytics. Prompt: "Call get_account_analytics for our Instagram and LinkedIn accounts for the last 30 days, presentation table, and tell me which format grew reach." Read calls run without approval prompts.

5. Reply to Google Business reviews. Prompt: "List new reviews on our Google Business Profile, draft replies with suggest_google_business_review_reply, and show me the options before posting." posterly generates on-brand reply choices without publishing until you say so; see the Google Business guide for the wider playbook.

6. Run a content calendar as a Scheduled Task. ChatGPT's Scheduled Tasks can run a recurring prompt every day or week, and tasks can use apps.²³ A Monday task that says "Check list_posts for gaps this week and propose three posts" is a decent editorial nudge. Task caps are Go 3, Plus 5, Business/Edu 10, Pro/Enterprise 15, and "Tasks cannot run more than once per hour", so leave the minute-level publish timing to posterly's queue.²³

7. Turn a long deliverable into a launch. ChatGPT Work is built for "longer, multi-step work and finished deliverables" and can start from a Project that holds your brand files and instructions.⁴ Ask Work to build the launch plan, then have it hand the individual posts to posterly.

8. Pre-flight everything. Before any create call, validate_post checks the full payload against each platform's rules without consuming quota, so a video that is too long for a platform gets flagged before it is queued.

ChatGPT vs Claude vs Grok Bot for MCP

The short version: Claude is the cheapest way to get write-capable MCP into a chat client, ChatGPT has the biggest audience and the most polished business admin controls, and Grok Bot is aimed at persistent agents on a team budget. Read the Claude guide and the Grok Bot guide for the mirror-image walkthroughs. Cursor, which SpaceX acquired in June 2026 according to VentureBeat, supports stdio, SSE, and Streamable HTTP through .cursor/mcp.json or ~/.cursor/mcp.json with OAuth for remote servers.²⁶²⁷ posterly also has dedicated setup pages for Hermes, OpenClaw, Grok Bot, Poke, and the posterly CLI, all riding on the same key and the same MCP endpoint.

Limits, security and gotchas

Plan gating is the number one support ticket. Free, Go, and Plus cannot add custom MCP servers at all; Pro is read-only; only Business and Enterprise/Edu can schedule through posterly from ChatGPT web.¹⁶ If you are on Plus and want a chat client with MCP writes today, the Claude guide covers a lower floor.

Prompt injection. OpenAI is direct about it: "Connecting to unsafe or untrusted MCP servers may increase exposure to security risks (including prompt injection). Only connect servers you trust", and "Some especially risky actions may be blocked instead of being presented for approval."¹⁶ The apps article adds that "Suspicious or hidden instructions in app content can also cause ChatGPT to ask for approval or block the request."¹³ The same warning applies on the API: "Prompt injection is an important security consideration ... especially true when you give the model access to MCP servers".¹⁸

Voice and training. Voice mode does not support apps, and on Free, Plus, Go, and Pro, app data may be used for training if "Improve the model for everyone" is switched on.¹³ Business and Enterprise workspaces are the safer home for client accounts.

OpenAI-built apps cannot write. OpenAI's first-party apps are "search-only today and do not support write actions. Use custom MCP apps for write/modify".¹⁶ That is why a custom app pointing at posterly, not a directory search app, is the path to scheduling.

posterly-side limits. Create-post calls are capped at 100 requests per hour per key, with separate higher limits for media writes and reads; create_posts_batch bundles up to 25 posts per request. Keys carry scopes such as posts:write and billing:read, and a key scoped to one workspace rejects calls for another. Hosted upload_media takes base64 up to about 5MB decoded; use upload_media_from_url or the stdio route for bigger assets. Rate-limited responses include retry_after, so ChatGPT can back off cleanly.

Retired features to stop citing. ChatGPT agent mode is gone ("ChatGPT agent is no longer available. Use ChatGPT Work for longer, multi-step tasks"), and its old 40 or 400 message caps are irrelevant.²⁸ Atlas is gone.⁵ The App directory is the Plugin directory.¹⁴ Context-window sizes by plan circulate on third-party sites, but we could not verify them on openai.com, so we do not quote them.

Best for, not best for

ChatGPT plus posterly is well-suited for

• Business and Enterprise teams that already live in ChatGPT and want scheduling inside the same workspace with admin approval controls
• Agencies that draft in Projects, finish in Work, and need one approval step before anything publishes
• Marketers who want recurring nudges via Scheduled Tasks without building a cron job
• Developers using Codex or the Responses API who want the same posterly tools in code and in chat
• Anyone who wants the largest ecosystem of first-party apps sitting next to their scheduler

ChatGPT is not the best fit for

• Free, Go, or Plus users who need write actions from a chat client; look at Claude instead
• Mobile-first workflows, since MCP apps are web only¹⁶
• Heavy local video pipelines, where a stdio client such as Claude Desktop, Codex, or Cursor is easier than JSON-only HTTP
• Always-on autonomous publishing, which is a better match for Grok Bot or a Hermes or OpenClaw runtime, with posterly still holding the queue

Final word

ChatGPT in 2026 is where most of your clients already are, and it now has a real tool layer: MCP-based apps, a plugin directory, developer mode for custom servers, Scheduled Tasks, and Work for long jobs. The gating is the only awkward part. If your workspace is Business or Enterprise, connecting a remote MCP server such as posterly is a ten-minute job and the approval prompts give you a sane audit trail. If you are on Plus, you are limited to reads until you upgrade or switch clients.

posterly meets ChatGPT on the transport it actually supports, hosted HTTP at https://www.poster.ly/api/mcp, with the same 18 platforms, the same scoped keys, and the same validation rules the dashboard uses. Start on the agents signup page or read the MCP page, then come back and ask ChatGPT to schedule your first week.

Footnotes