MCP and Skills

The desktop app supports the Model Context Protocol (MCP) for extensible tool integration and a skills system for adding specialized agent capabilities.

Model Context Protocol (MCP)

MCP is an open standard that lets AI agents interact with external tools and services. The desktop app includes built-in servers and supports installing additional ones.

Built-in Servers

These are included and automatically available:

Google Services Tools

When connected via OAuth, the Google server provides tools across:

• Gmail -- Read, send, search, manage emails
• Calendar -- Create, update, list events
• Drive -- Upload, download, search files
• Photos -- Access and organize photos
• Meet -- Schedule and manage meetings
• Tasks -- Create and manage task lists
• Contacts -- Search and manage contacts
• Sheets -- Read and write spreadsheet data
• Slides -- Create and modify presentations
• Docs -- Read and edit documents

Tools are filtered based on the permissions you grant during sign-in.

Web Search

When enabled in Settings > Search, the agent can search the web for current information. You can configure:

• Search mode: Auto (only for non-Claude backends), Always (all backends), or Manual (only when explicitly requested)
• Multiple providers: Add one or more from 13 supported providers and set their priority order
• Automatic failover: If the primary provider fails, the next one in priority order is tried

Supported providers include Tavily, Exa, Brave Search, Perplexity, You.com, Serper, Google CSE, DuckDuckGo (free), Jina (free), SearXNG (self-hosted, free), and more. Three providers (DuckDuckGo, Jina, SearXNG) work without an API key.

Installing MCP Servers

From the Marketplace

1. Go to Settings > MCP
2. Browse the Marketplace tab
3. Click Install on any server you want
4. The server is configured and ready to use

Available presets include:

• context7 -- Up-to-date library documentation
• sequential-thinking -- Enhanced reasoning chains
• filesystem -- File system operations
• github -- GitHub repository management
• playwright -- Browser automation
• slack -- Slack messaging

Cloud-Authenticated Servers

Some MCP servers connect securely to cloud services:

• Notion -- Access Notion workspaces
• Figma -- Design file access
• Granola -- Meeting notes

These use industry-standard OAuth for secure authentication.

DesignMode External MCP Starters

DesignMode also includes a starter catalog for design-focused MCP servers. The catalog covers Figma context, design token bridges, shadcn/ui snippets, Storybook metadata, 21st.dev Magic, Mermaid, AntV charts, Excalidraw-style architecture sketches, Photopea, ImageSorcery, Higgsfield, and Pollinations.

Some starters use OAuth, some use API keys, and some work without authentication but may be rate-limited. When you disconnect an external server, the app removes saved authorization data from its encrypted token store and stops sending that server's tools.

Slack App Home MCP Catalog

If you use the Slack bot, the Slack app's Home tab includes a quick-add catalog for personal MCP tools. You can add hosted connectors such as GitHub, Notion, Linear, Atlassian, and Sentry by pasting a provider token. Neumar encrypts the token and loads that MCP server only for your Slack-originated agent runs.

Admins can choose whether user-added MCP servers are available immediately, require review, or are hidden for a bot.

Managing MCP Servers During a Task

You can add, toggle, or reconnect MCP servers while a task is running -- without stopping the agent:

• Add a server -- Bring in a new MCP server mid-task
• Toggle a server -- Enable or disable a server without restarting
• Reconnect -- If a server loses its connection, reconnect it instantly
• View status -- See the health of all active servers

These changes apply only to the current task session and don't affect your saved MCP configuration.

Cross-Backend Compatibility

MCP tools work across all AI backends. Claude supports MCP natively, while other backends (like Gemini Local and OpenAI-Compatible) use an automatic compatibility layer. You don't need to configure anything -- the app handles the translation seamlessly.

Advanced: Manual Configuration

For advanced users, MCP servers can also be configured manually via JSON files:

• Shared configuration with Claude Code
• App-specific server configuration

Supported connection types:

• Local -- Servers running on your machine
• Remote -- Servers accessed via URL
• Cloud-authenticated -- Servers requiring OAuth sign-in

Skills and Plugins

Skills are reusable instruction sets that extend what the agent can do. They work like specialized "modes" the agent can activate. Skills are distributed inside plugins — installable bundles that can also include commands, hooks, and MCP servers.

What a Skill Looks Like

Each skill is a simple text file with a header and instructions:

---
name: my-skill
description: What this skill does
version: 1.0.0
---

Instructions for the agent when this skill is active...

Where Skills Come From

The desktop app loads skills from multiple places, in order of priority:

When two plugins share a name, the higher-priority source wins. Skill names are namespaced as pluginName:skillName so two plugins can ship a lint skill without colliding.

Browsing and Installing Plugins

Open Library in the sidebar. The page has these tabs:

To install a plugin:

1. Open Library > Marketplace
2. Search by name or description
3. Click a card to see the full manifest, requested permissions, and signature status
4. Click Install and confirm
5. The plugin appears immediately under Library > Plugins

To toggle or remove an installed plugin:

1. Open Library > Plugins
2. Click Enable / Disable to turn the plugin on or off without uninstalling
3. Click Uninstall to remove the plugin from disk

Signature Badges

Plugins can be cryptographically signed by their publisher. The install dialog shows one of three states:

For unsigned plugins, you must explicitly check the acknowledgement box before the Install button is enabled.

Configuring Remote Marketplaces

By default the Marketplace tab pulls from the Anthropic Official Plugins registry. You can point the app at additional registries — for example, a curated registry your team maintains on GitHub.

Each registry is a public URL serving a marketplace.json file. Registries are fetched in parallel, validated, and cached for 15 minutes. Partial failures don't break the tab — whatever loads is shown alongside an error message for whatever didn't.

Marketplace Safety

Plugin marketplace links are checked before the app loads them. The app blocks private network addresses, cloud metadata addresses, unsafe redirects, and malformed registry files.

Marketplace plugins also require a sandbox that is eligible for untrusted work. If only a reduced or no-isolation sandbox is available, the app will not silently fall back to running marketplace code directly on your computer.

Creating Your Own Plugin

Use the bundled CLI to scaffold a new plugin from a template:

pnpm plugin:new my-plugin
pnpm plugin:new my-plugin --template with-script
pnpm plugin:new my-plugin --template with-mcp --dir ~/work/plugins

Templates available:

After scaffolding, edit SKILL.md and install the plugin from disk via Library > Marketplace (or from the command line).

How Skills Work

When a plugin is installed and enabled, each of its skills becomes available to the agent. The agent reads the skill's instructions when it decides the skill applies to your request. Skills are non-destructive — they add capabilities without removing existing ones.

Skills in Agent Profiles

Agent profiles can bundle default skills. When you select a profile on the Home page, its skills are automatically activated alongside any globally enabled skills. This lets you create purpose-built agents with exactly the right skill set.

Restricting Skills

You can choose whether a profile has access to all installed skills or only a specific set:

• All Allowed (default) -- The agent can use any installed skill, plus all built-in capabilities like media generation and speech
• Restricted -- The agent can only use the skills you explicitly select. If you don't select any, the agent runs with the narrowest possible tool set

To configure skill restrictions:

1. Open the profile's Detail page
2. Go to the Tools tab
3. Toggle between All Allowed and Restrict
4. When restricted, check the skills you want to enable
5. Save the profile

If a previously selected skill has been uninstalled, it appears with a "remove" option so you can clean up stale entries.

Skill restrictions carry over to bot channels -- if a Slack, Telegram, Discord, or Lark bot uses a restricted profile, the bot's conversations respect the same skill limitations.

Skill Selector

The skill selector in the chat input groups skills by category with pinned skills shown at top. You can filter skills by name or trigger keyword.

Learn More

• Agent Runtimes -- Detect and manage local agent CLIs
• Google Workspace Tools -- OAuth-backed Gmail, Calendar, Drive, Docs, Sheets, and Slides tools
• Web Search -- Current-information search providers and failover
• Voice and Speech -- Speech synthesis and transcription providers
• Agent System -- How agents use tools during execution
• Desktop Application -- Overview and setup