@vantageos/mcp-agent-composer
v1.1.0
Published
MCP server: compose AI agents as Role + Persona + Framework + Skills. 12 roles, 10 personas, 16 framework refs. 5 typed tools. Bilingual FR+EN. stdio, no API key.
Maintainers
Readme
@vantageos/mcp-agent-composer
MCP server that composes an AI agent by mixing Role + Persona + Framework + Skills as a structured primitive. Bilingual FR+EN by design.
Version: 1.0.0 — GA (2026-04-26)
About
Most AI agents are defined by a single paragraph someone typed in a hurry. This server turns agent design into a structured, repeatable operation.
@vantageos/mcp-agent-composer gives any MCP-compatible client a typed composition primitive: pick a Role (what the agent knows), a Persona (how it communicates), an optional Framework (how it reasons), and custom Skills (what it can do). The result is a validated, locale-switched agent definition — as a system prompt, a JSON definition, or a Markdown card — generated consistently, session after session.
What it does
Five tools, each with a clear job:
| Tool | What you get |
|---|---|
| list_roles | All 12 curated roles with category filter (technical, creative, analytical, operational, leadership). Bilingual name + one-liner per role. |
| list_personas | All 10 curated personas with axis filter (formality, energy, directness, domain_focus). Includes voice traits and a sample phrase per persona. |
| compose_agent | Pass role_id + persona_id + optional framework_id + optional skills + context. Receive a complete agent definition in your chosen format: system_prompt, json_definition, or markdown_card. |
| suggest_composition | Describe what you need the agent to accomplish. Get 1-3 scored Role + Persona + Framework combinations with rationale — without knowing the catalog upfront. |
| validate_composition | Pass any role/persona/framework/skills combination. Get a compatibility score (0-100), severity-tagged warnings, and concrete recommendations before you deploy. |
Who it is for
- AI developers iterating on sub-agent design for Claude Code, Cursor, Goose, or ChatGPT MCP
- AI agencies that define multiple agent configurations per client and need consistent, auditable outputs
- Technical leads who want agent composition to be a typed primitive in their pipeline, not a prose convention
Why this, not hand-prompted agents or a prompt library
A hand-prompted agent definition drifts between sessions and teammates. A prompt library gives you templates with no validation. This server gives you a validated composition with a compatibility score, locale-switched outputs, and a repeatable UUID-tagged result — with no prompt engineering required from the caller.
The framework_id parameter shares the same 16-framework catalog as @vantageos/mcp-frameworks for cross-server compose: call suggest_framework on one server, pass the returned ID directly to compose_agent on this one.
Quick Start
npx -y @vantageos/mcp-agent-composerAdd to mcp.json:
{
"mcpServers": {
"vantage-agent-composer": {
"command": "npx",
"args": ["-y", "@vantageos/mcp-agent-composer"]
}
}
}No API key. No account. Restart your MCP client and ask: "Compose an agent for debugging production incidents — direct, technical, using first-principles reasoning."
Examples
Sub-agent design: Call compose_agent with role_id: "technical-architect", persona_id: "direct-pragmatist", framework_id: "first-principles", context describing your system. Receive a system prompt you can drop into any orchestrator immediately.
Discovery first: Use suggest_composition with goal "help a non-technical founder prioritize a feature backlog". Get scored suggestions — likely product-manager + warm-mentor + eisenhower — with rationale, before touching compose_agent.
Pre-deployment check: Hand-built a composition? Run validate_composition first. A provocative-challenger persona paired with a warm-mentor role scores low on compatibility — the server tells you why and what to swap.
Doctrine Flexibilite — Phase 1 / Phase 2
Phase 1 (current): stdio transport, local install, no API key, no remote server. Catalog data is fully offline — no LLM call at list/get time. Phase 2 (planned): HTTP transport for remote deployments + Pro tier with ALLOWED_ROLES scoping + VantageRegistry Skills lookup integration. Auth via Polar.sh. Activated based on adoption signal.
MIT License — Author: ElPi Corp / Laurent Perello — Source: github.com/elpiarthera/vantage-agent-composer-mcp
Installation
npx -y @vantageos/mcp-agent-composerConfiguration
Claude Desktop / Claude Code
{
"mcpServers": {
"vantage-agent-composer": {
"command": "npx",
"args": ["-y", "@vantageos/mcp-agent-composer"]
}
}
}Cursor
See examples/cursor.json.
Tools
| Name | Description |
|---|---|
| list_roles | List all 12 roles with one-line descriptions, optional category filter. |
| list_personas | List all 10 personas with voice traits and sample phrases. |
| compose_agent | Compose a complete agent definition (system_prompt / json_definition / markdown_card). |
| suggest_composition | Recommend the best 1-3 Role + Persona + Framework combos for a goal. |
| validate_composition | Validate compatibility and surface persona-role clashes. |
Catalogs
- 12 roles:
technical-architect,senior-developer,qa-engineer,data-analyst,product-manager,creative-director,copywriter,researcher,business-strategist,operations-manager,tech-lead,executive-coach. - 10 personas:
socratic-questioner,direct-pragmatist,warm-mentor,formal-academic,playful-creative,concise-executive,detailed-explainer,provocative-challenger,empathetic-listener,precise-analyst. - 16 framework IDs (mirrored from
@vantageos/mcp-frameworks):design-thinking,lean-startup,swot,okr,mece,first-principles,5-whys,eisenhower,raci,ooda,bcg-matrix,porter-5-forces,pareto,hofstede,cynefin,mckinsey-7s.
Examples
See examples/ directory.
Authentication
None in v1.0. Public local MCP server (stdio). No API key required.
The ALLOWED_ROLES environment variable for RBAC is deferred to Phase 2 — there is no authentication or authorization layer in v1.0.
Troubleshooting
- Server doesn't start — verify Node >= 20 (
node --version). - Tool not discovered — restart your MCP client after editing config.
- Invalid role_id / persona_id / framework_id — check the catalogs above (kebab-case enums).
- Locale not switching — pass
locale: "fr"explicitly in tool inputs. - Need French docs — see
README.fr.md.
License & Attribution
Author : ElPi Corp / Laurent Perello License : MIT Source : https://github.com/elpiarthera/vantage-agent-composer-mcp
