Skip to content

Extensions

Add skills, subagents, and MCP servers to an assistant.

Available on
  • Desktop
  • Web Portal

The Extensions tab on an assistant is where you plug in extra capabilities: ready-made skills the agent can invoke, custom subagents the harness can hand work off to, and MCP servers that expose new tools. Anything you add here is scoped to this assistant — not to your other assistants, and not to the project the assistant happens to be working in.

There’s no marketplace today; install is bring-your-own. Files come from your local filesystem, an inline editor, or (for Claude Code) auto-discovered from your home directory.

The Extensions tab is the only place Glueprint manages extensions. There’s no equivalent surface on projects — a Claude Code “project-scoped MCP” (the projects.<path>.mcpServers block in ~/.claude.json) is a Claude Code concept, not a Glueprint one, and applies filesystem-wide outside Glueprint’s UI. See Scopes on the MCP server reference for the full picture.

How to open it

  1. Open the assistant.
  2. Click Settings.
  3. Pick the Extensions tab.

The tab has three sections: Skills, Subagents, and MCP Servers.

Skills

A skill is a SKILL.md folder the harness will auto-discover when a user query matches its description. The harness reads the description, decides whether the skill is relevant, and (if so) reads the body for instructions.

Adding a skill

Click Add skill. A modal appears with two modes:

  • Inline — create the skill in place: pick a name (lowercase, hyphens, 1-64 chars), enter a description (this is what the harness sees when deciding whether to invoke), and write the body as Markdown.
  • Import — pick an existing SKILL.md folder from disk. Glueprint copies it into the assistant’s extension store.

Save. The skill appears in the Skills list.

Managing a skill

Each row shows the name, the description, a source pill (Inline or Imported), and a set of controls:

  • Toggle to enable / disable. Disabling preserves the canonical file in the store but stops it from being fanned out to the harness.
  • Edit — update the description or change enabled state.
  • Open folder — reveal the skill’s folder in your OS file browser so you can edit the body in any editor.
  • Re-sync — re-parse the skill’s frontmatter and refresh the display.
  • Remove — delete the skill with confirmation.

Subagents

A subagent is a Markdown agent file (.md) with optional frontmatter that lets the main harness hand sub-tasks off to a specialized agent. Frontmatter fields control which model the subagent uses, the list of tools it has access to, and a UI color.

Adding a subagent

Click Add subagent. The modal mirrors the Skills flow:

  • Inline — pick a name and description, then write the body as Markdown (with optional frontmatter at the top).
  • Import — pick a .md file from disk.

Save. The subagent appears in the Subagents list.

Managing a subagent

Each row shows the name, description, and a row of pills indicating any frontmatter the subagent declared — model override, tool count, custom color, and source (inline vs. imported). The same controls available on skills are available here: toggle, edit, open file, re-sync, remove.

Shadowing

If an inherited subagent (see below) and an assistant-level subagent share a name, the assistant-level one wins. This is intentional — it lets you override a globally-inherited subagent for a single assistant without removing it elsewhere.

MCP servers

MCP servers expose tools and resources to your assistant. The Extensions tab supports three transports.

Adding an MCP server

Click Add MCP server. The form depends on the transport you pick:

  • Stdio (a local process) — you supply a command, optional args (one per line), and optional env vars (key/value rows).
  • HTTP — you supply a URL and optional header rows.
  • SSE — same shape as HTTP.

All transports have a name field (a slug; reserved names that start with glueprint are blocked) and an Enabled toggle. A Show literal values toggle reveals environment variables and header values that are masked by default.

Environment variable expansion: write ${SOMETHING} and Glueprint substitutes the value from your shell environment when launching the server. Codex passes these literally (no expansion).

Harness differences

  • Claude Code — supports all three transports.
  • Codex — skips HTTP and SSE servers; only Stdio is honored. The UI flags this with a warning.
  • Gemini CLI — support matrix varies by version; the UI surfaces the actual capability.

Editing or removing

Edit any field except the name (renaming requires delete + re-add in the current release). The Enabled toggle stops the server from launching without losing its configuration.

Glueprint’s own MCP server

Glueprint ships its own built-in MCP server. It’s always available to every assistant — and to every human-initiated session — and cannot be removed or disabled from this tab. See The built-in MCP server for the tool reference and the assistant-mode filtering rules.

Inherited extensions (Claude Code only)

When the underlying agent is Claude Code, Glueprint also auto-discovers extensions you’ve already installed at the user level:

  • Skills from ~/.claude/skills/.
  • Subagents from ~/.claude/agents/ (and any Claude Code plugins).
  • MCP servers from the top-level mcpServers block in ~/.claude.json.

Inherited items appear in an Inherited subsection of each section with a “from <path>” label. They’re read-only at the assistant level — you can Open file or Refresh them, but to toggle, remove, or rename one, edit the file at its source.

Not shown here: project-scoped MCPs from projects.<path>.mcpServers in ~/.claude.json. Those are intentionally excluded from this view because they’re filesystem-scoped, not user-scoped, and applied by Claude Code directly. They’ll still run in any session whose working directory matches — just not through Glueprint’s UI.

PreviousWorkbench
NextAssistant governance