# Connecting an AI agent

Source: https://docs.tightly.io/help/ask/connect-an-ai-agent
Reviewed: 2026-10-03

## What it is

An AI agent is an assistant that runs outside Tightly, such as Claude, ChatGPT, Cursor or VS Code, and reads your organisation through Tightly's MCP server. MCP is the open standard these assistants use to connect to a service. Connected, the agent reads what your own seat reads, as you, and every call passes the same checks a press on the page passes. It can also ask Ask Tightly a question and read Tightly's own answer.

Connecting an agent is sold with Essentials+ and Pro. On Lite the page an agent opens refuses, in the plan's own sentence:

> Tightly Lite does not include connecting an AI agent. It is sold with Essentials+.

## The address to give your agent

Every agent connects to one address:

```
https://mcp.app.tightly.io/mcp
```

An organisation whose home is the EU region can use `https://mcp.eu.tightly.io/mcp`. The first time the agent reaches Tightly it opens your browser at [Connect an AI agent](../access/mcp-authorise), where you sign in, pick the organisation and choose its reach. An agent set up with an older address ending in /sse needs this one instead.

## Claude

On claude.ai or Claude Desktop, open **Customize**, then **Connectors**, press **Add** and **Add custom connector**, name it Tightly and paste the address. Leave sign-in on **Use Claude's published identity (Recommended)**. On a Team or Enterprise plan an owner adds it once under **Organization settings**, **Connectors**, and each member presses **Connect**. Turn it on in a conversation from the **+** button, **Connectors**.

In Claude Code, add it once, then type /mcp and sign in when the browser opens:

```
claude mcp add --transport http tightly https://mcp.app.tightly.io/mcp
```

## ChatGPT

Your own connections need a ChatGPT plan with developer mode. Turn on **Developer mode** under **Settings**, **Security and login**; some accounts show it under **Settings**, **Connectors**, **Advanced**. Then create an app: name it Tightly, paste the address, choose **OAuth**, save and connect. ChatGPT asks you before any tool that writes.

## Cursor and VS Code

Cursor reads mcp.json in the .cursor folder of your home folder or of a project:

```json
{ "mcpServers": { "tightly": { "url": "https://mcp.app.tightly.io/mcp" } } }
```

VS Code reads mcp.json in a project's .vscode folder, or run **MCP: Add Server** from the Command Palette:

```json
{ "servers": { "tightly": { "type": "http", "url": "https://mcp.app.tightly.io/mcp" } } }
```

Both open your browser to sign in the first time a Tightly tool is called.

## Read, or Read and act

**Read** lets the agent read what your seat reads and nothing else; no tool on Read writes. **Read and act** adds the act tools: the agent may propose adding lines to the Bench, drafting a purchase order or keeping an answer, and your own assistant asks you before each one, as the consent page says:

> It asks you before every act, and never signs off, releases or closes a commitment.

Read and act is offered only where the agent asked to act and your organisation allows it. An agent whose assistant cannot ask you is offered no act at all. A grant lasts 24 hours at a time, renewed for up to 30 days.

For an unattended integration, an admin can create a key on [API keys](../settings/api-keys) and give it to the agent in place of signing in. A key reads only the resources ticked on it, and it never acts or asks Ask Tightly. A key made in the EU region is used at `https://mcp.eu.tightly.io/mcp`.

## Ending an agent's access

Every agent that reads as you stands on [Profile](../settings/profile), under Your agents, with the day it was granted, when it was last used and when it last acted. An admin also sees every assistant connected to the organisation and can end any of them for everyone. **End access** stops one at once:

> This agent stops now; nothing it did is undone here.

The agent finds out on its very next call.

## Limits

A signed-in agent may make 120 calls a minute and propose 20 acts an hour. Over either, the call is answered with the wait, for example:

> This agent has asked 120 times this minute; try again in {wait}.

A key spends its own allowance, 600 reads and 120 writes a minute. What Tightly keeps is the consent page's promise:

> Every call is kept for 90 days, never what it read, and Tightly trains no model on your data.

## When it does not connect

- **The agent lists no Tightly tools.** Check the address ends in /mcp and that the assistant connects over streamable HTTP.
- **It reads but never acts.** It was granted Read, or its assistant cannot ask you. Reconnect it and choose Read and act, or make the change on the page.
- **The consent page refuses.** Its sentence says why: your plan, your organisation's list of approved agents, or a Tightly Connect seat.
- **The sign-in lapsed.** A request lasts 10 minutes; start again from the agent.
- **A figure looks old.** The agent reads Tightly live, so it shows what Tightly shows after the last sync.

## Glossary line

An AI agent is an assistant outside Tightly that reads your organisation as you through Tightly's MCP server, on a grant you choose and can end at any time.