# Getting started with AI

Comind.work lets you connect AI assistants to your workspace data, so you can query, analyze, and update records using natural language - through the AI tool you already use.

## AI capabilities overview[​](#ai-capabilities-overview "Direct link to AI capabilities overview")

| Feature                             | Description                                                                     |
| ----------------------------------- | ------------------------------------------------------------------------------- |
| [MCP tools](/ai/mcp-tools.md)       | Connect external AI assistants (Claude, ChatGPT, etc.) to your Comind.work data |
| [AI Assistant](/ai/ai-assistant.md) | In-app AI chat for workspace queries                                            |

The rest of this page focuses on MCP, which is the fastest way to get AI working with your data.

## What is MCP?[​](#what-is-mcp "Direct link to What is MCP?")

The **[Model Context Protocol](https://modelcontextprotocol.io/)** (MCP) is an open standard introduced by Anthropic in 2024. It defines how AI assistants connect to external systems to discover available tools, call them with structured parameters, and receive structured results - instead of copying data in and out of prompts.

Most major AI clients now speak MCP, including Claude Desktop, Claude Code, the claude.ai web app, Cursor, VS Code, and ChatGPT. An MCP connection is per-client and per-user: your AI tool talks to Comind.work on your behalf, using your credentials and your permissions.

Comind.work runs an MCP server at `https://mcp.comind.work/<your-company>`. Connecting any MCP-compatible client to it exposes your workspaces, apps, and records as tools the AI can call during a conversation.

## What you can do with MCP[​](#what-you-can-do-with-mcp "Direct link to What you can do with MCP")

Once connected, you can ask your AI tool questions like:

* **Query data** - "Show me all open tickets in the SUPPORT workspace assigned to me"
* **Aggregate and analyze** - "How many tasks were completed last month, broken down by team?"
* **Search** - "Find records mentioning 'budget approval'"
* **View history** - "What changes were made to TASK123 last week, and by whom?"
* **Read attachments** - "What does the screenshot in TICKET456 show?" (the AI can view images and read documents attached to records)
* **Create and update** - "Create a task in the IT workspace titled 'Rotate SSL certs' assigned to John", or "Change the priority of IT/TASK456 to high"

The AI picks the right tool automatically based on your question. You do not need to know which MCP tool corresponds to which action.

## Pick your client[​](#pick-your-client "Direct link to Pick your client")

Setup is different for each AI tool. Pick yours and follow the step-by-step guide:

|                          | Your AI tool                                        | Setup guide                                                                                                                                                                                                                                                                                                                                     |
| ------------------------ | --------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| ![](/img/ai/claude.svg)  | Claude Desktop, Claude Code (CLI), or claude.ai web | [Connect Claude](/ai/connect-claude.md)                                                                                                                                                                                                                                                                                                         |
| ![](/img/ai/chatgpt.svg) | ChatGPT (Plus/Pro/Team/Enterprise, developer mode)  | [Connect ChatGPT](/ai/connect-chatgpt.md)                                                                                                                                                                                                                                                                                                       |
| ![](/img/ai/cursor.svg)  | Cursor                                              | [Connect Cursor](/ai/connect-cursor.md)                                                                                                                                                                                                                                                                                                         |
| ![](/img/ai/vscode.svg)  | VS Code                                             | [Connect VS Code](/ai/connect-vscode.md)                                                                                                                                                                                                                                                                                                        |
|                          | Other MCP-compatible client                         | Use the server URL `https://mcp.comind.work/<your-company>` and the [authentication](#authentication) method below. See the [tool reference](/ai/mcp-tools.md) for what each tool does, and the [MCP spec guide on connecting to remote servers](https://modelcontextprotocol.io/docs/develop/connect-remote-servers) for client-agnostic setup |

Once connected, see the [MCP tools reference](/ai/mcp-tools.md) for the full list of tools, filter syntax, and limits.

## How it works[​](#how-it-works "Direct link to How it works")

<!-- -->

Three things to notice about this flow:

* The **AI client** translates your natural-language question into a structured tool call. It never sees your password or token directly - it calls the MCP server using the session you authorized.
* The **MCP server** enforces your user permissions before touching any data. If you cannot see a record through the UI, the AI cannot see it either.
* The **response** uses a compact tab-separated format (not JSON) with field descriptions and record slugs, so the AI understands the data's context without wasting tokens.

## Authentication[​](#authentication "Direct link to Authentication")

OAuth is the default - your client opens a browser window for you to sign in to Comind.work on first use. A long-lived API token is also available for clients that do not implement OAuth, or for automation. See [Authentication](/ai/authentication.md) for both setups, including the consent screen and example configs.

## Permissions and data handling[​](#permissions-and-data-handling "Direct link to Permissions and data handling")

* **Authentication** - every MCP call runs in the context of a user, either through an OAuth session or a manual API token tied to that user's account.
* **Permissions** - all queries respect workspace-level and app-level access controls. The AI sees exactly what you see in the UI.
* **Data boundaries** - the AI provider (Anthropic, OpenAI, etc.) sees the prompts you type and the Comind.work records returned by the tool calls. The Comind.work MCP server does not forward data anywhere beyond responding to the client that called it.
* **Token management** - rotate manual tokens periodically and revoke when no longer needed. OAuth sessions expire automatically.
* **Scope of access** - administrators can restrict MCP access using whitelist and blacklist patterns. Patterns can target a workspace (e.g. `HR`), an app across all workspaces (e.g. `*/INVOICE`), or a specific workspace-app pair (e.g. `CRM/QUOTE`). A workspace or app that is hidden from MCP will not appear in any tool call, even for users who otherwise have access to it.

Before connecting AI to a workspace with sensitive data, review your organization's data-sharing policies and the terms of your AI provider.

## FAQ[​](#faq "Direct link to FAQ")

### Does MCP cost extra?[​](#does-mcp-cost-extra "Direct link to Does MCP cost extra?")

No - Comind.work does not charge for MCP access beyond your normal subscription. You pay your AI provider (Anthropic, OpenAI, etc.) for the tokens consumed during conversations, at the rates defined by their plan.

Write operations are rate-limited per user. Read operations are not rate-limited, but individual tool responses are capped at 100,000 characters - use filters to narrow large result sets.

### Can the AI delete my data?[​](#can-the-ai-delete-my-data "Direct link to Can the AI delete my data?")

The MCP server exposes tools to **create** and **update** records, but not a general delete tool. Writes respect your user permissions: if you cannot perform an action through the UI, the AI cannot either. Permanent destructive actions (archiving, deleting) still require going through the UI.

### Which Claude model should I pick?[​](#which-claude-model-should-i-pick "Direct link to Which Claude model should I pick?")

If your client lets you choose:

* **Haiku** - the cheapest option, fine for simple lookups and short queries.
* **Sonnet** - fast and cost-effective for asking questions about data, quick aggregations, and routine updates.
* **Opus** - more reliable for multi-step work, app development, and anything requiring careful reasoning across many records. Higher cost per token.

Rule of thumb: Sonnet for asking questions, Opus for building or modifying apps, Haiku for simple lookups.

### Where is my data sent when I query?[​](#where-is-my-data-sent-when-i-query "Direct link to Where is my data sent when I query?")

The AI provider sees the prompt you type plus the Comind.work records returned by the tool calls your AI makes. The Comind.work MCP server runs within your company's Comind.work environment and does not forward data anywhere else - it responds only to the client that called it.

### Can administrators see what I ask the AI?[​](#can-administrators-see-what-i-ask-the-ai "Direct link to Can administrators see what I ask the AI?")

MCP calls are logged server-side like any other API call, so a workspace administrator can review which MCP tools you used and what records were returned through the admin activity log. The AI conversation itself (what you typed, how the AI responded) stays on your local machine and is governed by your AI provider's policies, not Comind.work's.

### What if the AI hallucinates a record?[​](#what-if-the-ai-hallucinates-a-record "Direct link to What if the AI hallucinates a record?")

Every record reference comes with a slug like `SUPPORT/TICKET456`. Before acting on what the AI says - especially for writes - open the record in Comind.work to confirm it exists and matches the description. If a slug does not resolve, treat the response as unreliable and ask again with more specific filters.

## AI tool compatibility[​](#ai-tool-compatibility "Direct link to AI tool compatibility")

|                          | AI tool           | MCP support   | Notes                                                                                                    |
| ------------------------ | ----------------- | ------------- | -------------------------------------------------------------------------------------------------------- |
| ![](/img/ai/claude.svg)  | Claude Desktop    | Full          | Recommended for Comind.work workflows - see [Connect Claude](/ai/connect-claude.md)                      |
| ![](/img/ai/claude.svg)  | Claude Code (CLI) | Full          | Configured via `.mcp.json` or `claude mcp add` - see [Connect Claude](/ai/connect-claude.md#claude-code) |
| ![](/img/ai/claude.svg)  | claude.ai (web)   | Full          | OAuth only, plan-gated - see [Connect Claude](/ai/connect-claude.md#claudeai-web)                        |
| ![](/img/ai/cursor.svg)  | Cursor            | Full          | Per-project or user-level config - see [Connect Cursor](/ai/connect-cursor.md)                           |
| ![](/img/ai/vscode.svg)  | VS Code           | Full          | Native MCP support or via Copilot/Cline/Continue - see [Connect VS Code](/ai/connect-vscode.md)          |
| ![](/img/ai/chatgpt.svg) | ChatGPT           | Limited       | Paid plans only, developer mode required - see [Connect ChatGPT](/ai/connect-chatgpt.md)                 |
|                          | Gemini            | Not available | Google supports MCP only in developer tools, not in gemini.google.com or voice mode                      |

## Next steps[​](#next-steps "Direct link to Next steps")

* [Connect Claude](/ai/connect-claude.md), [Connect ChatGPT](/ai/connect-chatgpt.md), [Connect Cursor](/ai/connect-cursor.md), or [Connect VS Code](/ai/connect-vscode.md) - set up your AI tool
* [MCP tools reference](/ai/mcp-tools.md) - what each tool does, filter syntax, limits
* [AI Assistant](/ai/ai-assistant.md) - the in-app AI chat experience