mcp2cli

name

mcp2cli

description

MCP-to-CLI lazy graph bridge. Connects any MCP server, OpenAPI spec, or GraphQL endpoint as a CLI command at runtime — zero codegen, 96-99% fewer tokens than loading tool schemas. Use when you need to call any MCP server tool but want to avoid loading all tool schemas into context. Supports lazy discovery: only load the server you need, when you need it. Triggers: "call MCP", "use supabase tool", "invoke vibe graph", "run watcher", any cross-server orchestration, or when token budget is tight.

MCP2CLI — Lazy Graph Bridge

The Emerald Tablet Law: Tools are vessels. They should be summoned only when the work demands them. Never load all MCP servers at once. The graph knows what is needed. Ask the graph.

---

CORE PRINCIPLE

This skill implements lazy graph loading for MCP tools. Each MCP server is a node in the Vibe Graph. Call only the node you need. Load its schema on demand. Release it after use.

MCP Registry (full graph)
    ├── gitkraken ── git/branch/diff/push
    ├── chrome    ── navigate/click/screenshot/fill
    ├── playwright── test/snapshot/evaluate
    ├── supabase  ── query/insert/select/rpc
    ├── higgsfield── video/animate/generate
    ├── kling     ── image-to-video/text-to-video
    └── worlds    ── create/deploy/voice-command
                          │
              summon only what you need
                          │
                    mcp2cli --mcp <url> <tool> <args>

---

INSTALL (one-time, idempotent)

# Preferred — no install, always latest
uvx mcp2cli --help

# Or install globally via uv
uv tool install mcp2cli

Windows note: ensure uv is on PATH. Install via winget install astral-sh.uv or pip install uv.

---

THE SERVER REGISTRY

Each MCP server the Kupuri OS uses. Load only when needed:

ID	Server URL	Purpose	Load when
gitkraken	stdio npx gitkraken-mcp	Git ops, PR, issues	Code changes needed
chrome	stdio npx mcp-chrome-bridge	Browser automation	Web scraping, testing UI
playwright	stdio npx @playwright/mcp	E2E testing	Automated testing
supabase	https://mcp.supabase.com/sse	DB queries, RPC, storage	Data ops
hostinger	stdio npx hostinger-api-mcp	Hostinger VPS, hosting, DNS	VPS / hosting tasks
higgsfield	https://api.higgsfield.ai/mcp	AI video	Media generation
kling	https://api.kling.com/mcp	Video synthesis	Video tasks
worlds	https://api.worlds.xyz/mcp	3D worlds	Immersive output

---

WORKFLOW (lazy load pattern)

Step 1 — Identify the minimum server needed

Task: "Query Supabase for sphere memory"
→ Server needed: supabase ONLY
→ Skip: gitkraken, chrome, playwright, etc.

Step 2 — Discover tools on that server (lazy schema load)

# List all tools on server (only loads this server's schema)
uvx mcp2cli --mcp https://mcp.supabase.com/sse --auth-header "Authorization:Bearer $SUPABASE_ANON_KEY" --list

# Search tools by keyword (even cheaper — returns names only)
uvx mcp2cli --mcp https://mcp.supabase.com/sse --search "query"

Step 3 — Call the tool directly

# Call a specific tool
uvx mcp2cli --mcp https://mcp.supabase.com/sse \
  --auth-header "Authorization:Bearer $SUPABASE_ANON_KEY" \
  execute_sql --query "SELECT * FROM vibe_nodes WHERE agent_id = 'alex' LIMIT 10"

Step 4 — Release. Do not cache schema in context.

---

STDIO SERVERS (local)

For MCP servers that run as local processes (gitkraken, playwright, chrome):

# List tools on a stdio-mode MCP server
uvx mcp2cli --mcp-stdio "npx @playwright/mcp" --list

# Call a tool
uvx mcp2cli --mcp-stdio "npx @playwright/mcp" \
  screenshot --url "https://dashboard-agent-swarm-eight.vercel.app" --output "./screenshot.png"

---

GRAPH ROUTING RULES

Follow these to keep token budget under control:

1. Never load more than 1 MCP server schema per task
2. Use --search before --list — search returns names only (20x cheaper than full schema)
3. Cache nothing in context — call mcp2cli fresh each time
4. If a built-in VS Code MCP tool covers it, use that instead of mcp2cli
5. For multi-server workflows, chain calls sequentially — never in parallel unless you parsed the output

---

EMERALD TABLET INTEGRATION

This skill is a Vessel (Tablet 4 — On Form and Substance).

When orchestrating Sphere agents that need cross-system data:

SYNTHIA needs → Supabase vibe_nodes + Git status + Screenshot
                     │                   │              │
              mcp2cli supabase    mcp2cli gitkraken   mcp2cli chrome
                   (sequential — one at a time)

Post results to Vibe Graph after each call:

# After getting data, POST to vibe graph
curl -X POST https://dashboard-agent-swarm-eight.vercel.app/api/vibe \
  -H "Content-Type: application/json" \
  -d '{"kind":"ingest","agentId":"synthia","nodeKind":"tool_result","label":"supabase-query","content":"..."}'

---

TOKEN BUDGET GUIDE

Operation	Token cost	Use when
--search "keyword"	~50 tokens	Finding tool name
--list	~500-2000 tokens	Need full schema
Direct tool call (known name)	~100 tokens	Tool name already known
Full MCP server in context	~5000-20000 tokens	NEVER — this is what we avoid

---

QUICK REFERENCE

# Install
uv tool install mcp2cli

# Discover
uvx mcp2cli --mcp <url> --search "<keyword>"

# List
uvx mcp2cli --mcp <url> --list

# Call
uvx mcp2cli --mcp <url> <tool-name> --<param> <value>

# Auth
uvx mcp2cli --mcp <url> --auth-header "Authorization:Bearer $TOKEN" --list

# Stdio server
uvx mcp2cli --mcp-stdio "<command>" --list
uvx mcp2cli --mcp-stdio "<command>" <tool-name> --<param> <value>

Hostinger uses the same lazy pattern with HOSTINGER_API_TOKEN in the environment:

HOSTINGER_API_TOKEN=*** uvx mcp2cli --mcp-stdio "npx hostinger-api-mcp" --list
HOSTINGER_API_TOKEN=*** uvx mcp2cli --mcp-stdio "npx hostinger-api-mcp" --search "virtual machine"
HOSTINGER_API_TOKEN=*** uvx mcp2cli --mcp-stdio "npx hostinger-api-mcp" list-virtual-machines