Connecting ChatGPT and ChatGPT classic

⚠ First: there are two apps

OpenAI installs two apps together: ChatGPT and ChatGPT classic. They now attach the same way: install AiAkiv Memory from the plugin directory inside the app (listed 2026-09-20). For most people that is all. Two cases use another method. To pin a project per folder you use an API key, on the ChatGPT app (§2). To use app features (asking the AI to do app work such as console management) you connect Codex separately with the install prompt or connect a folder with an API key (§6).

What older material called Codex is today's ChatGPT. That's also why the config file is .codex/config.toml. The install prompt in §6 is pasted into a Codex chat, such as the Codex CLI.

At a glance

• Install the plugin once and it applies account-wide, visible on the web and in both apps.
• The two setups can sit side by side. The folder side attaches under a different name, AiAkiv-<project name>, and you use that name to say which one to use → §4
• Even so, using only one is recommended → §5
• The official ChatGPT plugin does not support app features. To use them, connect Codex separately with the install prompt or connect a folder with an API key → §6

Where saving goes differs too. The plugin install and the Codex connection made with the install prompt follow Main from the console; a folder's key is pinned to that project (Teams and projects).

1. Install the plugin (both apps)

Search for it inside the app. There is no address to paste.

1. Open Settings → Plugins → Browse directory. The Plugins item in the sidebar opens the same place.
2. Type AiAkiv Memory in the search box and click the result.
3. Click Install plugin and sign in with your www.aiakiv.com account. One install applies account-wide, so ChatGPT classic sees it too.
4. Turn it on per conversation. Plugins are enabled per conversation. Under Plugins below a new chat's composer, turn AiAkiv Memory on and call it with @AiAkiv Memory. The Try now button shown right after installing opens a chat with it already on. Skip this and no tools appear in a new chat at all.

The plugin install is usually enough. Since AiAkiv was listed in OpenAI's plugin directory on 2026-09-20, there is no address to enter by hand if you only save, search and create cards. A connection made by registering the address yourself is only for using app features → §6.

Saving goes to Main. A plugin install has no mechanism to pin a project, so it lands in the Main project set in the console. The check-target tool shows binding as "main". To use different projects per folder, go to §2.

The official ChatGPT plugin does not support app features. The official AiAkiv Memory plugin in the ChatGPT plugin directory does not support app features (asking the AI to do app work such as console management). This connection does not have the app action tool, run_aiakiv_app_action. Saving memories, searching, reading linked teams, graph queries and creating cards work as usual. For console work such as teams, invitations, projects and managing cards, either do it in the web console yourself, or connect with one of the two methods in §6 and ask the AI → The app action tool.

2. ChatGPT: pin a project per folder with an API key

Installed per §1, the whole account follows Main. Use this method when you want different folders to save into different projects. Sign-in (OAuth) folder pinning (?project= in the connection address, or the X-K2G-Project header) does not attach on this app, so the one setup that pins a folder is a project-pinned API key.

With a key, the authentication step disappears entirely. Authentication is already inside the key, so there's no sign-in window and no browser (Authentication).

1. Issue that project's API key in the console (shown once, right after issuing). Use a project key here. A dynamic key attaches too, but its save location moves with Main, which defeats the point of pinning to a folder and makes it harder to tell where something was saved when you run both setups per §4. Use dynamic keys elsewhere (Authentication).

2. Put the following in the working folder's .codex/config.toml. The file location is the same for the app, the CLI, and the IDE extension.

   [mcp_servers.AiAkiv-myproject]
   url = "https://mcp.aiakiv.com/mcp"
   http_headers = { "Authorization" = "Bearer <your key>" }
   

3. Restart the app.

The name is AiAkiv-<project name>. AiAkiv-myproject in the example is the folder connection name for a project called myproject. Copy it from the console and it comes with your project's name already in it. Name it AiAkiv and it collides with §1's plugin install, so you can no longer tell them apart. The naming rule is in §4.

In the ChatGPT desktop app both connections show up. The desktop app reads the folder's .codex/config.toml, so it sees the folder connection (AiAkiv-myproject, pinned to this project) and §1's account-wide connection (the AiAkiv plugin, which follows Main) at the same time. The names differ, so name the folder side, as in "save through AiAkiv-myproject", and it works on this project (§3, §4). The Codex CLI works through the folder connection: run codex in that folder, and answer yes when it asks whether to trust the folder, or it won't read the config.

You can also get it all at once as a folder ZIP. On the console's project card, open This project's MCP config → Project-pinned API key and click Download folder ZIP. You get a folder holding the config files (.mcp.json, .gemini/settings.json, .codex/config.toml) and the AiAkiv skills. Unzip it and put the folder where you work. The skills for Claude are in .claude/skills/; the ones for Codex and Gemini are in .agents/skills/. If you haven't issued a key on that screen, the button issues one and puts it in the ZIP. Once you have the ZIP, clicking 'Issue a static key for this project' revokes the key inside it (a project has only one live key).

The project is pinned into the key. The coordinates are frozen inside it, so this folder doesn't move when you change Main in the console. The check-target tool shows binding as "api-key".

⚠ The key is your ID card. Leaking it means handing over access to your memory wholesale: don't commit config.toml. The full minimum is in Key hygiene in Authentication.

3. Both setups meet inside one app

In the ChatGPT app, both setups can be visible at once.

• The plugin install (§1) attaches to the account, so installing it once makes it visible in this app too.
• The folder's key (§2) is read when you start work in that folder.

So in the same app you end up with chat going through the plugin and folder work going through the folder's key. That isn't a problem in itself: but it does start to get confusing which one saved where. So to save into the folder project, name the folder-side connection (§4).

4. Running both: the folder connection has a different name

§1 and §2 can be used simultaneously. Having both is normal, and nothing blocks it. You tell them apart by name.

The global and folder MCP server names differ.

	Name	Where saves go
Global: plugin install	AiAkiv	Follows Main
Folder: .codex/config.toml	AiAkiv-<project name> (e.g. AiAkiv-k2g)	Pinned to that project

If both are AiAkiv, there's no way to say which one to use. The tool names collide too, so the AI picks whichever it likes at the time and you only learn where a save went afterwards. With different names you can say "save through AiAkiv-k2g."

How the folder connection is named. The config the console copies for you, and the folder ZIP, name the folder connection AiAkiv-<project name>. This is the same for Claude, Gemini and ChatGPT. The name goes into the tool names the AI calls (mcp__<connection name>__<tool>), and tool names may only contain English letters, digits, - and _. So sometimes the project name can't go in as is.

• Spaces and symbols become -. Project My Project → AiAkiv-My-Project. Runs of -/_ collapse into one, and a -/_ at either end of the name is dropped.
• Letters that aren't English (Korean, Japanese and so on) are dropped, and a short code taken from the project id is added. If almost everything is dropped, only the code remains (e.g. AiAkiv-3f9a1c). If only part is dropped, the code goes after what's left (e.g. K2G-연구 → AiAkiv-K2G-3f9a1c). The code differs per project.
• Names that are too long get cut. A connection name is at most 26 characters. A cut name also gets the short code at the end.

Where you can, name projects with English letters, digits, - and _ only. If a name has letters that aren't English, the connection name turns into a code that's hard to recognize. In that case you may change the server name inside the config files (.mcp.json, .gemini/settings.json, .codex/config.toml) to an English name that's easy to recognize (e.g. AiAkiv-research). The new name should also start with AiAkiv-, use only English letters, digits, - and _, and stay within 26 characters. In Claude Code you must also change the name inside enabledMcpjsonServers in .claude/settings.json to match. If the two names differ, the folder connection won't turn on by itself.

Claude is different. Even with different names the AI can use the account connector. Still, leave the account connector on (turning it off removes it from the web and the mobile app too) and check where each save went → When the account connector and a folder connection are both attached.

5. Recommended: use one, and if you use both, add a check

Use either §1 or §2, not both. With one setup there's no confusion about where saving goes.

If you must run both, keep two habits.

1. Name the side you want. When you give work, say which one by its §4 connection name.
2. Check the save target often. See §Checking below. ChatGPT tends to call get_save_target on its own fairly readily, so it follows a request to check.

Picking one switches sides (measured 2026-09-21). In a work (Codex) session with a folder key in place, plain "check the save target" answers from the folder project. Pick @AiAkiv-Memory (the plugin) in the composer of that same session and the same question ignores the folder and goes to the global Main. The mention is what decides. If you cannot remember which one you picked, check again before saving.

6. Using app features

App features means asking the AI to do app work such as console management: creating a team, inviting people, managing projects or managing the cards you made, from the conversation. The AI does this through the app action tool, run_aiakiv_app_action → The app action tool.

Roughly, the console app creates teams and sends invitations, handles team link invite codes, acceptance, pausing and resuming, shows the audit trail and the save target switch history, creates projects and edits personas, switches Main (the save target) and turns it off and on, lists and hides public targets, manages the cards you made, and shows which AI permissions are on. All of this can also be done by hand in the web console (the dashboard, app.aiakiv.com). Asking the AI is a convenience, and people who use only the official plugin can do every one of these in the web console.

The official ChatGPT plugin does not support app features. A connection installed per §1 is for saving memories, searching, reading linked teams, graph queries and creating cards. There are two ways to use app features in ChatGPT.

Method 1. Connect Codex (install prompt)

Why install it separately. The AiAkiv Memory plugin in the plugin directory does not have the app action tool, run_aiakiv_app_action. Parts of AiAkiv's app features don't fit OpenAI's plugin review policy. This tool is a general-purpose action tool, one tool that takes many kinds of work for many apps and runs them, and the plugin could not carry a tool like that. Yet all of AiAkiv's app features run through this tool: managing the console, writing a work in Contents, managing the cards you made, and other work where the AI controls things in detail. So the plugin still saves memories, searches, runs graph queries, reads linked teams and creates cards, but to use app features in Codex you install this connection separately.

What you get.

• App features: ask for console management, writing in Contents and managing the cards you made from the conversation, through run_aiakiv_app_action. Card creation (create_aiakiv_card) is there as well.
• Seven AiAkiv skills: aiakiv-onboarding, aiakiv-save-and-recall, aiakiv-graph-query, aiakiv-links, aiakiv-cards, aiakiv-console, aiakiv-contents. They go into ~/.agents/skills/<skill name>/, the same path as the skills in the console's folder ZIP.
• A sign-in connection: it attaches over OAuth, so there is no key to manage, and no developer mode is needed.
• Saving follows Main: the check-target tool shows binding as "main". To pin a project per folder you still use the API key in §2.

How.

1. Before pasting, deal with the existing plugin (see Naming rule below).
2. Paste the Codex install prompt into a Codex chat. Codex registers the address https://mcp.aiakiv.com/mcp?apps=all under the name AiAkiv, opens the sign-in window, installs the seven skills from the public repository github.com/AiAkiv/aiakiv-plugins, and writes an install record in the working folder.
3. When the sign-in window appears, sign in with your www.aiakiv.com account. A person does the sign-in.
4. Open a new session and paste the Codex verification prompt. If run_aiakiv_app_action is in the tool list, binding is "main" and the seven skills are read, you're done.

Naming rule.

• Remove the existing plugin where you can, and install this as AiAkiv. With both present you get two sets of tools with the same names, and the AI picks whichever it likes.
• If you keep the plugin, install this connection as AiAkivCodex. That's for when you also want the plugin in chatgpt.com web chats. Just change the MCP name in the install prompt. Then you can say which one to use, as in "save it through AiAkivCodex".

Where it applies. Measured on the Codex CLI (2026-10-13). The Codex desktop app and the IDE extension read the same settings according to OpenAI's documentation, but haven't been measured yet. This connection does not show up in chatgpt.com web chats; there, use the plugin from §1.

AI permissions. Console work in which the AI changes something only happens for what you turned on in Reference → AI permissions in the console. Reading is on from the start → Asking your AI to do console work.

Method 2. Connect a folder with an API key

This is the method in §2 as it is. Put a project-pinned API key in the working folder's .codex/config.toml and that folder is pinned to one project, and that connection has app features. For the AI to change anything in the console, check "Allow console management by AI (manage)" when you issue the key (Authentication).

Checking

Ask your AI before you start work.

aiakiv where are you saving right now?

The AI calls get_save_target and reports back. The binding value tells you which setup you're attached through.

If the project name carries a (Main) suffix, you're following the global Main.

When both connections are there, check by naming the folder connection: "call get_save_target through the AiAkiv-k2g connection". If binding is "api-key" (or "folder" for a folder setup attached with OAuth) and the project is right, you're attached through the folder connection. If it says "main", the account-wide connection answered.

Put aiakiv in the sentence (measured 2026-09-21). In a ChatGPT chat, asking only "check the save target" sometimes gets an answer with no tool call at all. Mixing the name in, as in "aiakiv check the save target", calls the tool right away. On the work (Codex) side, attached through a folder key, the tool was called without the name.

The account measured has a long history of using AiAkiv. On a freshly installed account the model's habits around picking tools may differ.

Every other value binding can take is in Confirming it attached in Folder binding.

Common problems

• I don't know which app I opened. Installing (§1) is the same on both. Pinning a folder (§2) works only on the app whose icon says ChatGPT.
• I added the folder config and it doesn't attach. Check whether you used the sign-in (OAuth) form (?project= in the address or the X-K2G-Project header). On this app the one setup that pins a folder is §2's key method.
• Saving goes to the wrong project. This happens when both setups are on and the names are identical. Rename the folder side per §4.
• No tools at all in a new chat. Plugins are enabled per conversation: step 4 in §1.
• I asked the AI for console work and it says this connection cannot do it. The official ChatGPT plugin does not support app features. Connect with one of the two methods in §6.
• I connected with the install prompt and run_aiakiv_app_action is missing. Run the Codex verification prompt in a new session. If the plugin is still there under the same name, the check may have gone through the plugin. Per the Naming rule in §6, remove the plugin or install as AiAkivCodex.
• The Codex connection doesn't show up in chatgpt.com web chats. That's expected. A connection made with the install prompt exists only on the Codex side. In web chats, use the plugin from §1.
• I removed the plugin, installed it again, and it won't attach. The old state is cached → Common problems.

Related

• Folder binding generally, and other clients → Folder binding
• Folder binding in Claude Desktop → Claude Desktop
• Authentication (OAuth and API keys) and key hygiene → Authentication
• The concepts of Main, projects, and the save target → Teams and projects
• Copying MCP config from the console → Console
• Adding app features in Codex → The Codex install prompt, The Codex verification prompt
• Which connections have the app action tool → The app action tool
• AI permissions and the console work you can ask for → Asking your AI to do console work