Best MCP Servers and Connectors for Meta Muse (2026)

Meta Muse includes a browser and a growing set of built-in connectors for services such as Notion, GitHub, Box, and Granola. It can also create Custom Connectors for services that expose an API or CLI. MCP follows a different route: consumer Muse does not currently document a way to add arbitrary MCP servers directly, while Muse Code supports MCP servers as external tools.

Meta also operates the Muse Connector Platform, where developers can submit integrations for review and distribution through Muse. That makes the right integration route dependent on both the product you are using and the job the agent needs to complete.

This guide compares six useful MCP servers and connector options by what they add to a Muse workflow, from web research and task management to repository access, automation, and smart-home control.

How MCP works in Muse

Muse supports built-in connectors and can also create Custom Connectors for third-party services that expose APIs or CLIs. Meta says these custom tools run inside Muse’s Secure VM, while credentials are kept in a separate secure store rather than exposed directly to the agent.

Consumer Muse does not currently expose a generic settings field where you can paste any third-party MCP server URL. That means an MCP-enabled service may still need to connect through Muse’s Custom Connector flow rather than through MCP directly.

We tested this with TinyFish. Muse created a Custom Connector and skill, then asked for a TinyFish API key through its Secure Credentials Store. Once connected, Muse could call TinyFish as a separate step for web tasks such as search, retrieval, and agentic browsing.

The tested setup was:

1. Ask Muse to use TinyFish for a web task.

2. Muse creates the Custom Connector and supporting skill.

3. Add agent.tinyfish.ai and your TinyFish API key in Muse’s Secure Credentials Store.

4. Muse can then route web tasks to TinyFish when needed.

How to evaluate an MCP server or connector for Muse

Start with the job the agent needs to complete rather than the number of tools a server exposes.

Task fit: Check whether the integration can perform the exact action required. Reading a workspace, searching the web, creating a task, and submitting a form are different capabilities.

Read vs. act: Determine whether the agent only needs information or must change something in an external system. Write access, browser interaction, purchases, form submissions, and other state-changing actions require tighter controls than read-only retrieval.

Authentication: Prefer documented authentication such as OAuth or narrowly scoped tokens, and avoid placing credentials directly in prompts or configuration that does not protect secrets.

Permission scope: Grant only the access required for the workflow. A research task should not receive administrative or destructive permissions simply because the connected account has them.

Deployment: Check whether the MCP server is remotely hosted or must run locally. Remote servers reduce infrastructure work, while self-hosted servers can provide greater control but require their own availability and security setup.

Output: Consider what the agent receives back. Structured results are useful when information must feed another workflow step, while application-specific connectors are often better when the agent needs to read or modify records inside one service.

6 Best MCP servers for Meta Muse

There is no single best MCP server for every Muse workflow. Notion, GitHub, Todoist, Home Assistant, Zapier, and TinyFish solve different problems, so the options below are organized by job fit rather than as a universal ranking.

For consumer Muse, check Meta’s built-in connectors first. Meta’s September 24 Connect recap confirmed Notion, Granola, GitHub, and Box among its work connectors, alongside additional commerce and travel integrations. When a built-in connector already covers the required service and action, it is usually the simplest route.

External MCP servers are most relevant in Muse Code. In consumer Muse, unsupported services can instead be accessed through a Custom Connector when they expose a suitable API or CLI.

1. TinyFish MCP server

Best for: Web research, page retrieval, and website workflows that require navigation or interaction.

TinyFish is MCP-native web infrastructure for AI agents, with Search, Fetch, Browser, and Web Agent available through one stack. In consumer Muse, however, our tested setup used Muse’s Custom Connector flow with a TinyFish API key rather than connecting the MCP URL directly.

TinyFish provides a hosted remote MCP server over Streamable HTTP with OAuth 2.1. Its MCP tools cover web search, content retrieval, and website automation, giving an MCP-compatible agent several ways to work with the web without routing every task through browser automation. Search and Fetch are free within published usage limits, while Web Agent and Browser use TinyFish’s Wallet-based metering.

What it adds

Personal agents such as Muse are increasingly able to act across connected applications, but the web is a different execution environment. A workflow may begin with research or retrieval and then encounter a site that requires multiple navigation steps, a form, authentication, or another interaction that cannot be completed through a simple connector.

TinyFish provides that web layer. An agent can start with Search, retrieve a known page with Fetch, and move to Web Agent only when completing the task requires interaction. This makes it useful for workflows that span research and web execution rather than staying inside a single SaaS application.

Muse route

Consumer Muse: We tested TinyFish through Muse’s Custom Connector flow using a TinyFish API key stored in Muse’s Secure Credentials Store. This setup uses TinyFish through its API rather than connecting the MCP endpoint directly.

Muse Code: Muse Code supports remote MCP servers over Streamable HTTP. TinyFish provides a hosted remote MCP server; see the TinyFish MCP integration docs for setup details.

Limitation

TinyFish is designed for web access and website workflows rather than replacing application-specific connectors. When Muse already has a native connector for a service such as Notion or GitHub, that connector is usually the more direct option for actions that stay within that application.

2. Notion MCP server

Best for: Searching, reading, creating, and updating workspace knowledge.

Notion’s hosted MCP server lets compatible AI clients work with pages and databases using the connected user’s existing Notion permissions. Notion warns that MCP tools can access everything the connected user can access, so account scope matters. Enterprise workspaces can use MCP Governance to approve specific AI apps and block unapproved clients.

Muse route

Consumer Muse now includes a built-in Notion connector, so that should be the first route for Muse users. Notion MCP is more relevant to Muse Code and other MCP-compatible clients.

Limitation

Notion MCP is designed for Notion content and actions. It does not replace open-web search, arbitrary URL retrieval, or website operation outside the workspace.

3. Todoist MCP server

Best for: Turning plans, meeting notes, or research findings into tasks and project updates.

Todoist operates an official MCP server at https://ai.todoist.net/mcp. Compatible AI clients can use it to read Todoist data, create and edit tasks and projects, and complete tasks. The connection uses OAuth and requests Todoist’s data:read_write scope. Users can review or revoke the resulting access token from Todoist’s integration settings.

Muse route

Todoist’s MCP endpoint can be used with an MCP-compatible client such as Muse Code. Meta’s September 24 Connect recap did not name Todoist among its work connectors. For consumer Muse, a Custom Connector against Todoist’s API is another route where the required operations are supported.

Limitation

Todoist is intentionally narrow. It is useful once a workflow needs to create or manage work, but it does not provide general web research, page retrieval, or access to systems outside Todoist.

4. Home Assistant MCP server

Best for: Querying smart-home state and controlling connected devices.

Home Assistant includes an official MCP Server integration exposed through Streamable HTTP at /api/mcp. It supports OAuth through Home Assistant’s authentication API, while clients that cannot use OAuth can connect with a long-lived access token. Administrators can control which Home Assistant capabilities are exposed to the MCP client.

Muse route

A reachable Home Assistant MCP endpoint can be connected to an MCP-compatible client such as Muse Code. Remote connections require a public Home Assistant URL, Home Assistant Cloud, or correctly configured networking such as a proxy or tunnel.

Limitation

The server is useful specifically for smart-home workflows. Device-control permissions should be granted more carefully than read-only state access.

5. GitHub MCP server

Best for: Repository context, issues, pull requests, and development workflows.

GitHub operates an official remote MCP server at https://api.githubcopilot.com/mcp/. Compatible clients can authenticate through OAuth or a personal access token, with access constrained by the approved OAuth scopes or PAT permissions.

The server can give an agent access to repository and development context without requiring a separate MCP deployment.

Muse route

Consumer Muse now includes a built-in GitHub connector, so use that route first when it covers the workflow. Muse Code can connect to GitHub’s remote MCP server directly.

Limitation

GitHub MCP is designed around GitHub. It is not a general web search or website-operation tool, and write permissions should be restricted when the workflow only needs repository context.

6. Zapier MCP

Best for: Taking actions across several business applications from one MCP connection.

Zapier MCP exposes actions from applications connected to a user’s Zapier account. Zapier currently supports more than 9,000 applications and 40,000 actions across its wider automation platform, allowing an AI client to work across tools such as email, calendars, CRMs, project-management systems, and other business applications.

Rather than loading every possible action into the agent’s context, Zapier can expose tools based on the applications already connected to the account. Users can also configure a fixed set of tools when tighter access is preferable.

Muse route

Zapier MCP is appropriate for Muse Code or another MCP-compatible client. Consumer Muse does not document arbitrary MCP endpoint setup, so its integration route depends on whether an appropriate built-in or Custom Connector is available.

Limitation

Zapier MCP uses the task allowance from the user’s existing Zapier plan. Each successful MCP tool call currently consumes two Zapier tasks, while failed calls do not consume tasks. It also depends on the actions Zapier exposes for each connected application, so it does not replace general web search, arbitrary-page retrieval, or website operation when no suitable app action exists.

Compare MCP servers and connectors for Meta Muse

Each option fits a different part of the Muse workflow. The table below compares what each integration is best suited for and the most practical route for using it with consumer Muse or Muse Code.

Comparison of MCP servers and connector options for Meta Muse by use case and integration route.

Security and permissions before you connect anything

Every external integration introduces another trust boundary. Use the smallest permission set that can complete the workflow and apply stricter controls when the agent can change data or act on a website.

• Verify the operator and endpoint. Prefer official servers where available and confirm how authentication, requests, and credentials are handled.
• Keep credentials out of prompts. Use OAuth, managed application connections, or appropriately protected tokens rather than exposing credentials directly to the model.
• Limit permissions. A workflow that only reads documents should not also receive delete or administrator access. Pay attention to inherited account permissions, particularly for services such as Notion.
• Treat actions differently from retrieval. Creating records, submitting forms, controlling devices, making purchases, or operating authenticated websites has a greater consequence than reading information. Require appropriate approvals for consequential actions.

Test new integrations with non-sensitive data before expanding their access.