Connect Instagram to ChatGPT with the HookSend MCP server
Supported, through a custom MCP connector added in ChatGPT’s developer mode and pointed at HookSend. ChatGPT sends you to a HookSend screen, you sign in and approve the workspace, and ChatGPT gets 63 tools for your campaigns, flows, inbox and stats. HookSend became an OAuth 2.1 authorization server for this on 1 October 2026.
One thing this page will not overstate: the flow is built to what OpenAI documents and covered end to end by our own tests, and no connection from ChatGPT’s own dialog has been confirmed yet, because we could not test against ChatGPT itself. Read what to do if your dialog behaves differently before you start. Written 30 September 2026, rewritten 1 October 2026.
How do I connect my Instagram account to ChatGPT?
Through a custom MCP connector in ChatGPT’s developer mode, pointed at HookSend. HookSendholds the Instagram connection, made on Instagram’s own authorisation screen through Meta’s official API, and offers 63 tools to whichever client connects. ChatGPT never touches Instagram directly and never sees your Instagram password.
Five steps, and the click path in them is the one OpenAI documents, read on 1 October 2026. Step four is the one that changed: there is no token to paste anywhere, because ChatGPT sends you to a HookSend page to sign in and approve a workspace instead. The section after the steps sets out what that sign-in actually does.
Step 1
Have a HookSend workspace that may connect an app
Sign in to HookSend on the 15-day trial or any paid plan. Connecting an AI client is not part of the Free plan, and a Free workspace is told so on the approval screen rather than after the connection appears to work. No MCP token is needed for ChatGPT — the sign-in replaces it.
Step 2
Turn developer mode on in ChatGPT
Open ChatGPT's Settings and turn Developer mode on. OpenAI's connect guide, read on 1 October 2026, puts the switch under Security and login; an older Help Center article puts it under Apps, then Advanced Settings. Look for whichever your account shows, and ask your workspace admin if the switch is missing.
Step 3
Add the connection and give it the address
Open ChatGPT's plugins or apps screen, press the plus button, give the connection a name and description, and enter https://hooksend.app/api/mcp as the server address. OpenAI's guide asks for the URL including the /mcp path, which that address already ends with.
Step 4
Choose OAuth, then approve it on HookSend's screen
Under Connection choose OAuth. ChatGPT registers itself with HookSend, then sends you to a HookSend page: sign in as normal, check the workspace named on it, read the three things the app will be able to do, and press Connect. ChatGPT is handed an access token that lasts an hour and refreshes itself.
Step 5
Read the tool list ChatGPT comes back with
ChatGPT lists the tools it found, 63 of them, and that list is the confirmation the connection works. Ask it what campaigns you have before asking it to change anything. Write actions confirm with you first — OpenAI's rule and HookSend's, independently of each other.
The address is the same for every client, and it already ends in the /mcppath OpenAI’s guide asks for.
https://hooksend.app/api/mcpWhat happens when ChatGPT asks me to sign in to HookSend?
A HookSend screen opens, and nothing is granted until you press the button on it. Sign in as normal, check the workspace the app may reach, read the three things it will be able to do, and press Connect. ChatGPT is handed an access token that lasts an hour, not your password and not a HookSend token.
Everything in the table below is a fact about HookSend’s own code as of 1 October 2026, not a description of the protocol in general. Two of the rows can be checked from outside without an account at all: fetch either discovery document and compare it with what is written here.
| Topic | What HookSend does, as of 1 October 2026 |
|---|---|
| How a client finds the flow | A request with no token, or with a token HookSend will not accept, is answered 401 with a WWW-Authenticate header naming the scope and pointing at the protected-resource document for the exact address the request arrived on. Fetching that document needs no account and no token. |
| The two discovery documents | Protected-resource metadata (RFC 9728) is served at /.well-known/oauth-protected-resource and at the path-inserted form ending /api/mcp, and names the one authorization server. Authorization-server metadata (RFC 8414) is served at /.well-known/oauth-authorization-server and carries the authorize, token, registration and revocation endpoints. |
| PKCE | Mandatory, and S256 only. The plain method is neither advertised nor accepted, and a request that omits the challenge method is refused rather than assumed to mean S256. A code exchange with the wrong verifier, or with none, issues nothing. |
| Registering the client | Dynamic client registration is open, and no client secret is ever issued: every client is public and authenticates with PKCE alone. A registration on its own reaches nothing — it buys the right to ask a signed-in person for consent. Return addresses must be https, or http on loopback, and are matched exactly. |
| Who approves it | A person, on a HookSend screen, every time. The ordinary HookSend login runs first, then a consent screen that names the app, names the workspace, lets you change it if you belong to more than one, and lists what the app will be able to do. Nothing is auto-approved and nothing is granted by the client alone. |
| How long the access lasts | An access token lasts one hour. The refresh token rotates on every use and expires thirty days after its last use, so an abandoned connection dies by itself. A replayed authorization code or a reused refresh token revokes the whole connection rather than just failing. |
| What the access covers | One scope, one workspace. There is no read-only variant, because a scope that said read-only while the tools behind it can write would be worse than one honest scope. The app cannot see your other workspaces, change your plan, invite anybody, or create HookSend tokens. |
| Which address the token works at | Only the one it was issued for. A token minted for the address below is refused at any other HookSend address, and the reverse, so connect on the address printed here and leave it alone afterwards. |
| Turning it off | HookSend settings lists the app under Connected apps, with the date it was connected and when it was last used, and a Disconnect button that stops it reaching the workspace straight away. Everything the app did stays in your activity log after it is gone. |
Will my Claude, Claude Code, Codex or Antigravity connection stop working?
No, and nothing about it needs redoing. The pasted-token path is unchanged: the same token, the same address, the same Authorization header, and every connection configured before 1 October 2026 keeps working untouched. OAuth was added beside that path rather than in place of it, and a working token never meets the new flow at all.
Worth saying because the temptation is to go and tidy something up. A token is bound to one workspace by the person who created it, which is a stronger binding than OAuth gives, and revoking it in settings still cuts the client off immediately. Claude, Claude Code, Codex and Antigravity each accept a header and are documented that way on the hub page and, for the one that reads a TOML file, on the Codex page.
Why is there no Connectors option in my ChatGPT settings?
Usually because OpenAI renamed it. OpenAI’s developer documentation calls the area Plugins and points at chatgpt.com/plugins, while OpenAI’s Help Center article on developer mode said Apps and connectors. The developer pages were read on 1 October 2026 and the Help Center article on 30 September 2026, so the label on your screen depends on which version of the interface your account is on.
Three other reasons account for most of the rest. Developer mode has to be switched on before a custom connector can be created, and OpenAI’s two pages put the switch in two different places — Settings, then Security and login on the newer one; Settings, then Apps, then Advanced Settings on the older. The Help Center says the feature is web only, so a phone will not show it. And in a managed workspace an administrator has to enable it before anybody in that workspace sees it at all.
Worth knowing that none of this is HookSend’s to fix, and none of it is a sign that something is wrong with your account. OpenAI is moving this area: the same documentation says custom GPTs are being retired and that a custom GPT’s actions do not carry through the migration, which is why a connector is the supported way to give ChatGPT a tool now.
What does OpenAI document about custom MCP connectors?
Every row below is a fact from OpenAI’s own pages — the authentication guide, the developer-mode guide, the connect guide and the Help Center article — and each row carries the date the page it came from was read. Two rows record OpenAI disagreeing with itself about menu names and plans, and the disagreement is printed rather than resolved, because a reader looking at one of the two versions is not helped by a page insisting on the other. Nothing here is HookSend’s rule, and OpenAI can change any of it without telling us.
| Topic | What OpenAI’s own pages said | Read on |
|---|---|---|
| What the area is called | OpenAI's developer documentation calls it Plugins and points at chatgpt.com/plugins. OpenAI's Help Center article on developer mode used Apps and connectors. Both wordings were live on the dates beside this row, so look for whichever your own account shows. | 1 October 2026, and 30 September 2026 for the Help Center |
| Turning developer mode on | OpenAI's connect guide gives Settings, then Security and login, then Developer mode. The older Help Center article gave Settings, then Apps, then Advanced Settings. In a managed workspace an admin has to enable it as well. | 1 October 2026, and 30 September 2026 for the Help Center |
| Transport | The connect guide asks for the MCP server URL including the /mcp path, and offers a Secure MCP Tunnel with a tunnel id as the alternative. A server running on your own machine needs that tunnel; ChatGPT does not reach a local process. | 1 October 2026 |
| Authentication | OAuth 2.1, No Authentication, or a mix of the two, where the mix means the handshake and the tool list go unauthenticated and the tools themselves do not. OpenAI's authentication guide rules out machine-to-machine grants and says ChatGPT cannot present custom API keys or customer-provided mTLS certificates. A static token in a header is not on the list, which is why HookSend now speaks OAuth. | 1 October 2026 |
| How the client gets registered | Three ways, in OpenAI's own order. Static credentials are used when the app's creator supplies them; Client ID Metadata Documents are called the preferred method where the authorization server advertises support; and dynamic registration is used when the creator chooses it or CIMD is unavailable. HookSend implements the third. | 1 October 2026 |
| What the server must publish | PKCE, provably. OpenAI's authentication guide requires an authorization server to publish code_challenge_methods_supported containing S256 so a client can confirm PKCE support from the metadata alone. HookSend publishes exactly that, and publishes nothing else in its place. | 1 October 2026 |
| Plans | The developer-mode guide names Pro, Plus, Business, Enterprise and Education, on the web. The Help Center article was narrower and said full MCP was for Business and Enterprise or Edu, with Pro limited to read and fetch. Free is not named on either page. | 1 October 2026, and 30 September 2026 for the Help Center |
| Mobile | Web only, in the Help Center article's own answer. The same article says agent mode will not use custom apps, and that deep research can but only for read and fetch actions. | 30 September 2026 |
| Write actions | The developer-mode guide says write actions require confirmation by default, that remembering an approval lasts for one conversation only, and that a tool without a read-only hint is treated as a write action. | 1 October 2026 |
| Changing the server later | A connection has to be reopened and refreshed to pick up new tools, and in a managed workspace ChatGPT uses a frozen snapshot until an admin refreshes it. Nobody is notified when the underlying server changes. | 30 September 2026 |
What if the ChatGPT connection dialog does not do what this page says?
Tell us what it asked for, and use another client in the meantime. The flow above is built to what OpenAI’s authentication guide documents and is covered end to end by HookSend’s own tests, and no connection from ChatGPT’s own dialog has been confirmed yet, because we could not test against ChatGPT itself. Two gaps are known and named below rather than left for you to find.
Client ID Metadata Documents
OpenAI’s authentication guide calls CIMD the preferred way to register a client where the authorization server supports it, and HookSend does not implement it. The same guide says a client registers dynamically when CIMD is unavailable, which is the path HookSend serves, and our metadata deliberately does not claim CIMD support — advertising a capability we do not have would send ChatGPT down a road that ends in a refusal it cannot recover from.
A dialog asking for a client id and secret
OpenAI documents static credentials as the first option, used when the app’s creator supplies them. HookSend has no screen that hands you a client id, because every client here is public and registers itself, so a dialog that insists on one is a dead end today rather than something to work around. Say so and it gets built; that is a small change, not a redesign.
Meanwhile the fallback is a real answer rather than a consolation. The identical address works in Claude, in Claude Code, in Codex and in Google Antigravity, each of which takes a HookSend token in a header, and HookSend does not know or care which client is calling. Nothing about your Instagram connection, your campaigns, your leads or your allowance changes with the choice, so a ChatGPT attempt that stalls costs you the ten minutes and nothing else.
What can ChatGPT actually do with my Instagram account?
Run the automation, not just report on it. Once connected, ChatGPT can create a comment-to-DM campaign on a reel, build a DM flow from a description, write a Comment Guard rule in plain words, reply inside an existing DM thread, and read your stats, leads and post performance — 63 tools, the same ones the dashboard sits on top of.
Four tool names, so the claim is checkable rather than atmospheric: campaigns_create makes a campaign, flows_create makes a DM flow, stats_overview reads the numbers and leads_recent reads the leads a campaign captured. Comparing that with the Instagram connectors that only pull Insights into a chat window is the subject of a page of its own, linked below.
Three limits are Meta’s rather than ours, and ChatGPT inherits all three. Meta allows one private reply per comment, ever, across every campaign and every tool. Meta allows a reply inside a DM thread for 24 hours from that person’s last message. Meta requires the connected account to be an Instagram professional account, so a personal account cannot be connected at all.
Where these ChatGPT steps come from
All four pages below are OpenAI’s, and every ChatGPT claim on this page was read from them on one of the two dates printed beside it. Two of them disagree about menu names and about which plans qualify, which is why both are listed rather than the one that suits us. OpenAI can change any of it without telling us, so if one of these now says something different, believe it rather than this page.
- Plugin authentication
The page that documents OAuth 2.1, rules out custom API keys and machine-to-machine grants, requires S256 to be published in the server's metadata, and sets out the three ways a client gets registered. The one to read before anything else.
- Developer mode
OpenAI's page for custom MCP connectors: the authentication mechanisms, when Client ID Metadata Documents and dynamic registration are each used, the plans it names, and the rule that write actions are confirmed by default.
- Connect to ChatGPT
The click path, button by button, including where the developer-mode switch sits, the plus button on the plugins screen, and the request for the server URL with the /mcp path on the end.
- Developer mode and MCP apps in ChatGPT
OpenAI's Help Center article, which uses the older Apps and connectors wording and gives a narrower list of plans than the developer docs do. Read on the earlier of the two dates on this page; it would not serve us on the later one.
Claims about HookSend’s own side have a different source: the code that serves the endpoint, and the test file that pins its behaviour. Where this page says a token lasts an hour or that a replayed code kills a connection, that is what the tests assert, not a promise about a roadmap.
Questions people ask before they connect it
- How do I connect my Instagram account to ChatGPT?
- Through a custom MCP connector added in ChatGPT's developer mode, pointed at HookSend, with OAuth as the connection method. HookSend holds the Instagram connection — made on Instagram's own authorisation screen, through Meta's official API — and exposes 63 tools to whatever client connects. ChatGPT sends you to a HookSend page to sign in and pick a workspace; it never touches Instagram directly and never sees your Instagram password.
- Does ChatGPT support OAuth for a custom MCP server?
- Yes, and OAuth is the mechanism OpenAI documents. OpenAI's plugin authentication guide, read on 1 October 2026, documents OAuth 2.1 for MCP servers, rules out machine-to-machine grants, and says ChatGPT cannot present custom API keys. It also requires an authorization server to publish code_challenge_methods_supported containing S256. HookSend became an OAuth 2.1 authorization server for its MCP endpoint on 1 October 2026 and publishes exactly that.
- Do I need a HookSend MCP token to use ChatGPT?
- No. A token is what Claude, Claude Code, Codex and Antigravity use, and ChatGPT's connector dialog has no field to paste one into, which is the reason OAuth exists here. With ChatGPT you sign in to HookSend instead, on a HookSend screen, and approve one workspace. Nothing is copied, nothing is stored in a configuration file, and you can disconnect the app again from HookSend settings.
- Will my Claude, Claude Code, Codex or Antigravity connection stop working?
- No, and nothing about it needs redoing. The pasted-token path is unchanged: the same token, the same address, the same Authorization header, and every connection configured before 1 October 2026 keeps working untouched. OAuth was added beside that path rather than in place of it, and a working token never meets the new flow at all. Revoking a token still stops it immediately, as before.
- Why is there no Connectors option in my ChatGPT settings?
- Three reasons, in the order worth checking. OpenAI has renamed the area: its developer documentation calls it Plugins where the Help Center said Apps and connectors. Developer mode has to be switched on before a custom connector can be created at all. And the Help Center says the feature is web only and not on every plan, so a phone or a Free account will not show it. In a managed workspace an admin has to enable it before it appears for anyone.
- Which ChatGPT plans can add a custom MCP connector?
- Paid plans, on the web, and OpenAI's own two pages do not agree on which. The developer-mode guide names Pro, Plus, Business, Enterprise and Education, read on 1 October 2026. The Help Center article, read on 30 September 2026, said full MCP was for Business and Enterprise or Edu, with Pro able to connect read and fetch tools. Free is not named on either page, so treat it as unsupported.
- What if the ChatGPT connection dialog does not do what this page says?
- Tell us what it asked for, and use another client in the meantime. No connection from ChatGPT's own dialog has been confirmed yet, because we could not test against ChatGPT itself, and two gaps are known: HookSend does not implement Client ID Metadata Documents, which OpenAI calls the preferred registration method where a server supports it, and there is no screen that hands you a client id if ChatGPT ever asks for static credentials. The same address works today in Claude, Claude Code, Codex and Antigravity with a token.
- Will ChatGPT change my Instagram account without asking me?
- No, and two separate mechanisms stop it. OpenAI's developer-mode guide says write actions require confirmation by default and that an approval is remembered for one conversation only. Independently, HookSend refuses the first call to each of 9 actions that cannot be undone and tells the client to come back and ask you, which is enforced by HookSend rather than by asking the model nicely. No HookSend tool deletes a campaign, a flow or a knowledge base at all.
- Can ChatGPT read my Instagram DMs?
- Only the conversations Meta exposes to the connected professional account, and only through HookSend. ChatGPT cannot produce an archive of everything ever sent to you, cannot read a thread nobody has written in, and cannot read anything from before the account was connected, because Meta's API does not offer any of those. Replying inside an existing thread is possible through inbox_send_reply, within the 24 hours Meta allows from that person's last message, and it is one of the actions that asks you first.
- Which HookSend plan do I need to use ChatGPT with my Instagram account?
- The 15-day trial or any paid plan, from $6 a month (₹99 in India). Connecting an AI client is not part of the Free plan, so a workspace that has finished its trial and moved to Free is stopped on the approval screen with the reason on it. Whatever ChatGPT does counts against the same monthly DM allowance and AI credits as the dashboard, so the connector is not a way around a plan limit.
Read next
The hub at /mcp lists all 63 tools and the per-client settings, and the longer explainer covers what MCP is from the beginning.
Connect it to ChatGPT, or to whichever client you already use
15 days of every feature, no card, and the connector is included. ChatGPT signs in and approves a workspace; Claude, Claude Code, Codex and Antigravity take a token in a header. One workspace serves all of them at once. Paid plans start at $6 a month (₹99 in India); the Free plan does not include the connector.
Start free