MCP servers

The Model Context Protocol is a standard external servers follow to expose tools to an AI agent. Riggr ships a native MCP client and merges those servers' tools with the tools your canvas nodes generate.

Practical result: Riggr's agent gets access to resources that don't live inside the canvas — a GitHub API, a filesystem outside the workspace, persistent memory, anything.

When you want an MCP

• You need an integration Riggr doesn't have a native node for (Linear, Notion, Stripe, Sentry).
• You want the same tools config shared between Riggr and Claude Desktop.
• You want the agent to have memory that persists across workspaces.

If the integration already exists as a node (Git, Docker, Database, etc.), prefer the node — it ships with UI and typed ports, MCP only gives you tools.

Adding a server

Two ways.

1. Through Settings (recommended)

Ctrl+Shift+P → Settings → MCP. Click Add server and fill in:

• Name — any identifier.
• Transport — stdio (default), http, or sse.
• Command + args (for stdio) — e.g. npx -y @modelcontextprotocol/server-github.
• Env vars — keys the server needs, like GITHUB_TOKEN.
• Enabled — toggle. Keeps in the list but paused.

Save. Riggr spawns the server on next restart (or click Reload).

2. Editing the file

Click Open config file in the MCP tab. It opens riggr-mcp.json (in <userData>) in your system editor. The format accepts both the Riggr-native shape and Claude Desktop's (mcpServers):

{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": { "GITHUB_TOKEN": "ghp_..." }
    },
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/me/Documents"]
    }
  }
}

After saving, Settings → MCP → Reload or restart the app.

Store presets

Riggr ships an in-app store with ready-made configurations:

• Filesystem — access to a specific directory.
• GitHub — repos, issues, PRs, code search.
• Fetch — pull content from arbitrary URLs.
• Memory — lightweight vector store for cross-session memory.
• PostgreSQL — query and schema for a remote Postgres.

In Settings → MCP, click Browse store, pick one, fill in the variables, install. The server lands in your config.

How the agent sees it

When you start an Agent Terminal:

1. Riggr collects tools from every active MCP server + tools generated by connected nodes.
2. Generates a temporary config file and passes it to the CLI agent (Claude Code, usually) via --mcp-config.
3. The agent calls tools with prefixed names — MCP tools become riggr_mcp_<server>_<tool>, node tools become riggr_<action>_<node>.

From the conversation side, it's all the same — you ask, the agent uses whatever's closest at hand.

Lifecycle

• Startup: enabled: true servers come up in parallel at app launch. A failure doesn't block the app.
• Crash detection: if the process dies, Riggr marks it as errored and shows the status in the MCP tab.
• Reload: Settings → MCP → Reload stops and restarts everything that's active.
• Shutdown: tree-kill on app close — makes sure child processes die with it.

Troubleshooting

• Server won't connect — open logs at Settings → MCP → View logs. Most errors are a missing env var or a command not in PATH.
• Tool doesn't show up for the agent — confirm the server is connected. Spin up a fresh Agent Terminal (the current one already captured the old list).
• Tool name conflict between servers — the server suffix resolves it, but if you want total control, rename a server in the config.