twist-ai
# Twist AI and MCP SDK
Library for connecting AI agents to Twist. Includes tools that can be integrated into LLMs,
enabling them to access and interact with a Twist workspace on the user's behalf.
These tools can be used both through an MCP server, or imported directly in other projects to
integrate them to your own AI conversational interfaces.
## Using tools
### 1. Add this repository as a dependency
```sh
npm install @doist/twist-ai
```
### 2. Import the tools and plug them to an AI
Here's an example using [Vercel's AI SDK](https://ai-sdk.dev/docs/ai-sdk-core/generating-text#streamtext).
```js
import { fetchInbox, reply, markDone } from '@doist/twist-ai'
import { streamText } from 'ai'
const result = streamText({
model: yourModel,
system: 'You are a helpful Twist assistant',
tools: {
fetchInbox,
reply,
markDone,
},
})
```
## Using as an MCP server
### Quick Start
You can run the MCP server directly with npx:
```bash
npx @doist/twist-ai
```
### Setup Guide
#### Claude Desktop
Add to your Claude Desktop configuration file (`claude_desktop_config.json`):
```json
{
"mcpServers": {
"twist": {
"command": "npx",
"args": ["-y", "@doist/twist-ai"],
"env": {
"TWIST_API_KEY": "your-twist-api-key-here"
}
}
}
}
```
#### Cursor
Create a configuration file:
- **Global:** `~/.cursor/mcp.json`
- **Project-specific:** `.cursor/mcp.json`
```json
{
"mcpServers": {
"twist": {
"command": "npx",
"args": ["-y", "@doist/twist-ai"],
"env": {
"TWIST_API_KEY": "your-twist-api-key-here"
}
}
}
}
```
Then enable the server in Cursor settings if prompted.
#### Claude Code (CLI)
```bash
claude mcp add twist npx @doist/twist-ai
```
Then set your API key:
```bash
export TWIST_API_KEY=your-twist-api-key-here
```
#### Visual Studio Code
1. Open Command Palette → MCP: Add Server
2. Configure the server:
```json
{
"servers": {
"twist": {
"command": "npx",
"args": ["-y", "@doist/twist-ai"],
"env": {
"TWIST_API_KEY": "your-twist-api-key-here"
}
}
}
}
```
### Getting your Twist API Key
1. Visit [https://twist.com/app_console](https://twist.com/app_console)
2. Create a new integration or use an existing one
3. Copy your API key
4. Add it to your MCP configuration as shown above
## Features
A key feature of this project is that tools can be reused, and are not written specifically for use in an MCP server. They can be hooked up as tools to other conversational AI interfaces (e.g. Vercel's AI SDK).
This project is in its early stages. Expect more and/or better tools soon.
Nevertheless, our goal is to provide a small set of tools that enable complete workflows, rather than just atomic actions, striking a balance between flexibility and efficiency for LLMs.
### Available Tools
- **userInfo** - Get information about the current user and their workspaces
- **fetchInbox** - Fetch threads and conversations from the inbox
- **loadThread** - Load a specific thread with its comments
- **loadConversation** - Load a specific conversation with its messages
- **searchContent** - Search across a workspace for threads, comments, and messages
- **createThread** - Create a new thread in a channel. Accepts an optional `displayInInbox` boolean (default `false`). When `true`, the thread is unarchived after creation so it appears in the author's Inbox. See also [TWIST_CREATE_THREAD_DISPLAY_IN_INBOX](#environment-variables).
- **reply** - Reply to threads or conversations
- **react** - Add reactions to threads, comments, conversations, or messages
- **markDone** - Mark threads or conversations as done (read and/or archived)
- **buildLink** - Build URLs to Twist resources
For more details on each tool, see the [src/tools](src/tools) directory.
## Environment Variables
| Variable | Default | Description |
| -------------------------------------- | ------------ | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `TWIST_API_KEY` | _(required)_ | Your Twist API key. |
| `TWIST_CREATE_THREAD_DISPLAY_IN_INBOX` | `false` | Set to `true` to unarchive every newly-created thread so it appears in the author's Inbox, without needing to pass `displayInInbox: true` on each call. **Only takes effect when running the MCP locally.** The remote/hosted MCP does not have this variable set and will use the per-call `displayInInbox` parameter only. |
## Dependencies
- MCP server using the official [@modelcontextprotocol/server](https://github.com/modelcontextprotocol/typescript-sdk?tab=readme-ov-file#installation)
- Twist TypeScript SDK [@doist/twist-sdk](https://github.com/Doist/twist-sdk-typescript)
## Local Development Setup
### Prerequisites
- Node.js 18 or higher
- npm
- A Twist account with API access
### Setup
1. Clone the repository:
```bash
git clone https://github.com/doist/twist-ai.git
cd twist-ai
```
2. Install dependencies:
```bash
npm install
```
3. Create a `.env` file with your Twist API key:
```bash
TWIST_API_KEY=your-twist-api-key-here
```
4. Build the project:
```bash
npm run build
```
### Development Commands
- `npm start` - Build and run the MCP inspector for testing
- `npm run dev` - Development mode with auto-rebuild and restart
- `npm test` - Run all tests
- `npm run type-check` - Run TypeScript type checking
- `npm run format:check` - Run linting and formatting checks
- `npm run format:fix` - Auto-fix linting and formatting issues
## Contributing
Contributions are welcome! Please ensure:
1. All tests pass (`npm test`)
2. Code is properly typed (`npm run type-check`)
3. Code passes linting and formatting checks (`npm run format:check`)
Use [Conventional Commits](https://www.conventionalcommits.org/) for commit messages:
- `feat:` for new features
- `fix:` for bug fixes
- `docs:` for documentation changes
- `test:` for test changes
- `chore:` for maintenance tasks
## License
MIT
TDQS
Scored across 19 tools
Tools are mostly distinct with clear scopes: load-thread vs load-conversation vs fetch-inbox vs search-content vs get-mentions all retrieve content but with specific purposes. The generic update-object and delete-object use targetType to differentiate, reducing ambiguity. Only slight overlap between get-users and user-info, but their purposes are different enough.
Naming follows a consistent kebab-case, verb-first pattern (e.g., get-users, load-thread, create-thread, mark-done). A few single-word verbs like reply, react, and away deviate from the compound style but still follow a verb-based convention. Overall, naming is predictable and consistent.
With 19 tools, the set is substantial but appropriate for a full-featured messaging platform API. Each tool addresses a distinct aspect (workspaces, channels, threads, conversations, reactions, search, etc.), and while a few are auxiliary (build-link), the count feels well-scoped for the domain.
The tool set covers core operations for threads, conversations, messages, reactions, user management, and search. However, there is no direct tool for creating or deleting channels, and no way to list threads in a channel without using search or inbox. Updating a conversation is also unsupported (update-object only handles threads, comments, and messages). These are notable gaps for complete lifecycle management.