> ## Documentation Index
> Fetch the complete documentation index at: https://docs.usesyndicate.org/llms.txt
> Use this file to discover all available pages before exploring further.

# MCP servers

> Connect agents to external tools and data with Model Context Protocol servers.

MCP (Model Context Protocol) servers give agents tools beyond the built-ins: GitHub, databases, internal APIs, anything with an MCP server. In Syndicate you install servers once, globally, then turn them on per agent.

## Install a server

Open **Settings, MCP**. Two paths:

### From the Catalog

The Catalog offers "One-click presets for common MCP servers. Ones that need credentials will ask for them first." Click **Install**, or **Configure** if the preset needs credentials: a modal collects them ("Secrets are stored locally and synced into Claude's MCP settings."), with a **Setup docs** link when available.

### Add a server manually

Under **Add a server**, choose a transport:

| Transport                 | Use when                           | You provide                                                       |
| ------------------------- | ---------------------------------- | ----------------------------------------------------------------- |
| **Local command** (stdio) | The server runs as a local process | Command (usually `npx`) and args (`-y package-name` plus options) |
| **Remote URL** (HTTP)     | The server is hosted               | An `https://` URL                                                 |
| **Remote SSE**            | The server streams over SSE        | An `https://` URL                                                 |

Give it a name and a short "What agents use it for" description, add any **Secrets** as key/value pairs (values are entered as passwords and encrypted; they never land in plain-text config), and click **Add server**. Syndicate validates the server with a live handshake before saving it.

## Manage installed servers

Each installed server row shows a health dot (green healthy with tool count, red failed with the error, gray unchecked), its transport badge (Local, HTTP, SSE), and actions:

* **Test**: re-runs the handshake and lists the tools the server exposes.
* **Show details**: the exact command or URL plus available tools.
* **Toggle**: enable or disable the server globally.
* **Remove**: uninstall, with confirmation.

## Turn servers on per agent

Installing a server does not give every agent access. In each agent's configuration panel, the **Tools (MCP)** section lists installed servers with per-agent toggles. Enable only what the agent's role needs.

After enabling, verify in chat: ask the agent to use the tool ("List my open GitHub PRs"). Under **Ask first** permissions you will get a Tool Request prompt; allow it and confirm the result.

<Tip>
  Scope tools to roles. A researcher with a web or GitHub server makes `/investigate` dramatically stronger; a DevOps agent with a CI server gets live pipeline visibility for `/deploy-review`. But there is rarely a reason for every agent on the team to carry every tool: fewer tools means better tool choices and a smaller blast radius.
</Tip>

## Security notes

* Secrets are encrypted with your OS keychain-backed storage and injected into the server process at launch; they are never written to settings files.
* MCP tool calls go through the same [permission flow](/reference/permissions-and-safety) as built-in tools, so "Ask first" agents still prompt you before using them.