# Connect your AI (MCP)

Connect the assistant you already use — ChatGPT, Claude, Claude Code, Codex or Cursor — to your Instasent project, so it works with your audience, campaigns and reports in plain language.

**Language:** en
**Audience:** platform
**TLDR:** Instasent runs an MCP server at https://develop-mcp.instasent.com/mcp. In a web assistant like ChatGPT or Claude you add it as a custom connector, log in and enable it per conversation. In a coding assistant like Claude Code, Codex or Cursor you paste one prompt and it sets everything up itself, including Instasent Skills (published as the Agent Skill `instasent`). Either way you accept the permissions as a package and pick a project, and from then on it reads your audience, campaigns and reports and prepares work with you. It never sends a message or activates anything — you confirm that in the dashboard — and you can revoke it at any time from Connected apps.
**Translation key:** platform.developers-and-apps.connect-your-ai
**Search keywords:** connect your ai, claude, claude code, chatgpt, cursor, codex, gemini cli, mcp, model context protocol, mcp server, custom connector, connector, connect an agent, ai agent, ai agents, agent, agent skills, skills, instasent-cdp, playbooks, oauth
**Related pages:** /platform/es/developers-and-apps/connect-your-ai, /platform/en/developers-and-apps, /platform/en/developers-and-apps/managing-connected-apps, /platform/en/developers-and-apps/third-party-data-access
**Docs index (every page):** https://docs.instasent.com/llms.txt
**This zone's index:** https://docs.instasent.com/platform/en/llms.txt
**This page:** https://docs.instasent.com/platform/en/developers-and-apps/connect-your-ai/ (HTML) · https://docs.instasent.com/platform/en/developers-and-apps/connect-your-ai.md (Markdown)
**Other language (es):** https://docs.instasent.com/platform/es/developers-and-apps/connect-your-ai.md

You can connect the AI assistant you already use so that it works directly with
your Instasent project. Instead of clicking through the dashboard and copying
figures out of it, you ask in your own words: *"how many customers ordered
before but not in the last three months?"*, *"draft a campaign for that
group"*.

It works because Instasent runs an **MCP server**. MCP (*Model Context
Protocol*) is the open standard assistants use to plug into other tools; in
practice it is the shared language that lets yours understand your project. The
model is yours — you connect it, and it keeps running wherever you already use
it.

## Which one are you using?

Setting it up is different depending on where your assistant lives, so start
here and follow only your half.

- [A web assistant](#web-assistants-chatgpt-and-claude) - ChatGPT, Claude and the like. You add Instasent as a connector from a settings screen — no terminal, nothing to install.
- [A coding assistant](#coding-assistants-claude-code-codex-cursor-and-gemini-cli) - Claude Code, Codex, Cursor, Gemini CLI. You paste one prompt and your assistant sets itself up.

## Web assistants: ChatGPT and Claude

#### 1. Give your assistant the address

Wherever it asks you to add an **MCP server** or a **custom connector**,
paste this URL:

```text
https://develop-mcp.instasent.com/mcp
```

That is the whole configuration: no keys, no files, nothing to install. The
name you give it is only so you recognise it in your own list — call it
whatever you like.

#### 2. Log in and approve

Your assistant sends you to an Instasent screen. You sign in, see who is
asking, **which permissions** and **for which project**, and approve. Your
password never reaches the assistant.

#### 3. Turn it on in the conversation

Most assistants keep connectors off by default in each new chat. Enable
Instasent from the conversation's tools or connectors menu, and start
asking.

> **Note**: Permissions are granted **as a package**: you accept or deny the whole set the
> assistant asks for, rather than ticking boxes one by one. It is how the
> connections you already use work, and it avoids the classic failure of leaving
> out one permission that then breaks half the answers.

#### Claude

Go to **Customize → Connectors**, choose **Add custom connector** and paste
`https://develop-mcp.instasent.com/mcp`. This shortcut opens that dialog with the fields already
filled in — you still confirm it:

[Add Instasent to Claude](https://claude.ai/customize/connectors?modal=add-custom-connector\&connectorName=Instasent\&connectorUrl=https%3A%2F%2Fdevelop-mcp.instasent.com%2Fmcp)

Then, in each conversation, enable Instasent from the **+ → Connectors**
menu. On a **Team or Enterprise** plan an owner can add it once for the
whole organization from **Organization settings → Connectors**, and members
then only press **Connect**. On the **Free** plan Claude currently allows a
single custom connector.

#### ChatGPT

Custom connectors live behind developer mode: **Settings → Apps &
Connectors → Advanced settings** to enable it, then **Settings →
Connectors → Create**, paste `https://develop-mcp.instasent.com/mcp` and choose **OAuth** as the
authentication method. Complete the Instasent login, and enable the
connector per conversation from the tools menu.

> **Tip**: Your assistant will most likely warn you that Instasent is an **unverified**
> application the first time. That is expected for connectors added by URL, and
> Instasent shows you the same thing from its side: you always see who is asking
> before you approve.

### Instasent Skills, in a web assistant

**Instasent Skills** are how Instasent's operating knowledge reaches your
assistant: how an audience filter is written so it doesn't quietly select
nobody, what a message really costs once its encoding is counted, which figures
are approximations and must not be quoted as exact.

**You already have them.** The connection serves them live, and your assistant
reaches for them on its own — there is nothing to install for it to work. What
follows is only for keeping a copy loaded permanently.

- [Download Instasent Skills](https://develop-mcp.instasent.com/.well-known/skills/instasent-skills.zip) - Everything in one file, published in the Agent Skills format assistants read. Downloading it needs no account and no credential.

In **Claude**, go to **Customize → Skills**, choose **Upload a skill** and pick
that file.

Other assistants add knowledge in their own ways, and those ways change often.
Rather than guess for you, ask yours — paste this:

```text
Instasent publishes its product knowledge — Instasent Skills — as an Agent
Skill, a folder of markdown files. I can download it as a zip from:
https://develop-mcp.instasent.com/.well-known/skills/instasent-skills.zip

Two things:

1. Can you load knowledge like that — a skill, a knowledge file, a project
   file, anything equivalent? If you can, tell me exactly where in your
   interface I upload it. If you can't, say so plainly and don't improvise
   a workaround.

2. Either way, commit this to memory for our future conversations: the same
   knowledge is available live from the Instasent connector, through its
   get_playbook tool. Before answering anything about Instasent — how an
   audience filter is written, what a message costs, what a figure includes
   — call get_playbook and answer from that, rather than from what you
   already believe about the platform. Save that instruction to your memory
   now.
```

> **Warning**: **Skills installed in one place stay in that place.** Uploading them to Claude
> does not put them in Claude Code, and installing them in Claude Code does not
> put them in Claude — the two don't share. If you use both, install in both.

## Coding assistants: Claude Code, Codex, Cursor and Gemini CLI

These can set themselves up. Paste this and let your assistant do the whole
thing — the connection and Instasent Skills, in one go:

```text
Set Instasent CDP up in this project. Run the commands yourself rather than
asking me to run them.

1. Check whether the Instasent MCP server is already configured, and add it
   if it isn't. It is a remote server speaking Streamable HTTP at
   https://develop-mcp.instasent.com/mcp — there are no keys, the URL is the whole configuration.
   In Claude Code that is:
     claude mcp add --transport http instasent https://develop-mcp.instasent.com/mcp
   In another assistant, add that same URL to your MCP config file.

2. Check whether Instasent Skills are already installed, and install them
   if they aren't:
     npx skills add https://staging-instasent-docs-nextjs.oscar-284.workers.dev

Then stop and tell me three things: what it covers, after reading the
SKILL.md you installed; what I have to do to finish signing in; and whether
this session has to be restarted before the Instasent tools show up.
```

That closing instruction is doing real work. **Signing in is yours** — it opens
a browser and only you can approve it — and most assistants only notice a new
MCP server on their next session. Asking your assistant to tell you both saves
you finding out by silence.

> **Note**: **Claude Code**, **Codex**, **Cursor** and **Gemini CLI** all read Instasent
> Skills from the folder the installer writes to — the first two verified by us,
> the other two documented by their makers. It knows where dozens of other
> assistants keep theirs too, and leaves the files where each one looks.

#### I'd rather do it by hand

Three steps, and you can stop after the first if you only want the
connection.

**1 · Add the MCP server.** In Claude Code:

```bash
claude mcp add --transport http instasent https://develop-mcp.instasent.com/mcp
```

In another assistant, add that same URL to whatever config file it uses for
MCP servers. Then sign in — in Claude Code, `claude mcp login instasent`.

**2 · Install Instasent Skills.** One command, and it finds your
assistant's skills folder on its own:

```bash
npx skills add https://staging-instasent-docs-nextjs.oscar-284.workers.dev
```

What it asks is **where**. *In the repository* travels with the code, so
everyone on the team — and any agent working on that codebase — gets it
without installing anything; that is the one you want as a team. *Just for
you* applies to every project you open, and `--global` picks it without
being asked.

**3 · Or skip the installer** and unzip the download above into the folder
your assistant reads skills from. For Claude Code:

```bash
unzip -o instasent-skills.zip -d ~/.claude/skills/
```

#### You already had it installed under the old name

Instasent Skills used to be called `instasent-cdp` and are now called
`instasent`. That name names the folder, so reinstalling **does not update
the one you had: a new one appears beside it**.

Keep both and your assistant reads both, with the old one frozen at the
version you installed that day. Delete it:

```bash
rm -rf .agents/skills/instasent-cdp
```

That is the path if you installed it *in the repository*. If you installed
it just for you it sits in your assistant's skills folder, which in Claude
Code is `~/.claude/skills/instasent-cdp`.

The same goes for the MCP connection if you follow the steps above as they
stand: you would end up with `instasent-cdp` and `instasent` pointing at the
same server, and the same tools twice over. The old name still works, so you
can keep it and change nothing, or drop it:

```bash
claude mcp remove instasent-cdp
```

## What a connected assistant does

It works with a **curated set of tools**, not with the whole API. In practice it
can look at your project the way you would and describe what it finds: who is in
your audience and how many match a condition, what you have been sending and how
it went, what your automations do, and what your project can send today. It can
also leave work prepared — a segment saved from the audience you just defined
together, a campaign written and left as a draft — and it can read this
documentation on its own, so it answers questions about Instasent without you
pasting links.

The exact list grows and changes with each release, so the honest way to think
about it is by shape: **it reads your project and prepares work; it does not
carry it out.**

### What it never does

> **Warning**: **No assistant sends.** There is no tool that sends a message, launches a
> campaign or starts an automation. It can leave everything ready — the segment
> saved, the campaign drafted — and you press send in the dashboard. That
> boundary does not depend on the permissions you granted: it is not in the
> surface your assistant sees, under any permission.

It doesn't import your contacts or export your list either. That lives in the
API, for integrations that carry their own credentials — see
[API tokens](/platform/en/developers-and-apps/api-tokens).

## Checking it works

The honest test isn't listing a directory, it's asking for something only a
connected assistant could answer:

> *"Recommend a winback segment for my store, and tell me how many contacts it
> would have."*

A good answer comes back with **a concrete proposal and a number from your
project** — customers who ordered before but not since a certain date, counted.
If instead you get general advice about winback campaigns and no number, it
isn't connected.

That question is a good test precisely because it needs both halves: your data
for the count, and Instasent Skills to build the segment the way this platform
actually stores that — recency lives on the contact, not in the events.

> **Warning**: **An installed skill is a photograph.** It is compiled from the live server in
> the same release, so it can never describe something that isn't there — but
> once it is on your disk, the platform keeps moving. **Where your installed copy
> and a live answer disagree, the live answer is right and the copy is old.**
> Updating is one command, worth running now and then:
> 
> ```bash
> npx skills update
> ```

## Staying in control

- The connection reaches **only the project you chose** and only the permissions
  you approved.
- Your project's **access type** applies on top of that, and can hold back or
  anonymize personal data: an assistant can end up able to count an audience
  without being able to see who is in it. See
  [Data access](/platform/en/developers-and-apps/third-party-data-access).
- Every connected assistant appears in
  [Connected apps](/platform/en/developers-and-apps/managing-connected-apps)
  with its permissions, and you can **disconnect it at any time**. It loses
  access immediately.

> **Tip**: Using an assistant that can't connect over MCP? It can still read these guides:
> point it at
> [https://staging-instasent-docs-nextjs.oscar-284.workers.dev/platform/en/llms.txt](https://staging-instasent-docs-nextjs.oscar-284.workers.dev/platform/en/llms.txt),
> a plain-text index it can pull single pages from.

Building your own app or agent on top of Instasent, for other customers to
install? That is the developer side:
[Connected apps (OAuth)](/developers/product-api/connected-apps-and-oauth).

---

This is one page of the Instasent documentation. For the complete machine-readable index of every guide and API reference, fetch https://docs.instasent.com/llms.txt — start there for full context.