MCP servers and skills

Connect MCP servers and give the agent skills.

MCP servers

Settings › MCP Servers lists every server with Global and Project switches, its status and its tools.

Add a server under mcp.servers in ~/.ngram/settings.json or the project's .ngram/settings.json (Open settings.json on that page):

"mcp.servers": {
  "github": {
    "url": "https://api.githubcopilot.com/mcp/",
    "headers": { "Authorization": "Bearer ${GITHUB_TOKEN}" }
  },
  "postgres": {
    "command": "npx",
    "args": ["-y", "@modelcontextprotocol/server-postgres", "${DATABASE_URL}"]
  }
}
KindFields
Localcommand, args, env, env_file
Remoteurl, headers. OAuth sign-in runs when the server asks for it. Add "type": "sse" for older SSE servers.
  • Servers start with your login shell's environment, so npx, uvx and docker are found.
  • ${NAME} and ${env:NAME} are filled from the environment, and ${workspaceFolder} from the open folder.
  • Values entered in Settings › MCP Servers are stored in the Keychain.
  • In Read and Plan modes, an MCP call asks unless an allow rule covers it. See Saved permissions.

Importing from other tools

ngram finds servers set up in Claude Code, Claude Desktop, Cursor, Codex, Gemini CLI, VS Code and Windsurf. Turn one on in Settings › MCP Servers or on the setup page and ngram copies it to your settings.

Skills

A skill is a SKILL.md file the agent loads when a request matches it.

ScopeLocation
Global~/.agents/skills/<name>/SKILL.md
Project.agents/skills/<name>/SKILL.md
---
name: release-notes
description: Write release notes from the commits since the last tag.
---
Steps the agent follows.
  • name is lowercase letters, digits and -, and matches the folder name.
  • disable_model_invocation: true means only you can run it.
  • Every skill is also a /name command in the composer.
  • Settings › Agents › Skills has New skill and Import from GitHub, and Copy share link on each skill.
  • Changes apply without a restart.