mcpx

One CLI for any MCP server -- built for humans and AI agents.

Transform MCP servers into ergonomic CLI commands. Register once, invoke anywhere.

Quick Start

# Install
npm install -g mcpx

# Register a server
mcpx add weather "npx @mcp/server-weather"

# Discover tools
mcpx list /weather
mcpx schema /weather get-forecast

# Invoke (agent-optimized)
mcpx /weather get-forecast --params '{"city": "Berlin"}'

# Invoke (human-friendly)
mcpx /weather get-forecast --city Berlin

Why mcpx?

For humans -- stop wiring MCP servers by hand. mcpx gives you --help on every tool, --dry-run to preview, and interactive mode for exploration.

For AI agents -- every response is structured JSON with semantic exit codes. Your agent gets a deterministic CLI interface with zero custom tooling.

Commands

| Command | Description | |---------|-------------| | mcpx add <alias> <cmd-or-url> | Register an MCP server (stdio or HTTP) | | mcpx remove <alias> | Remove a registered server | | mcpx servers | List registered servers | | mcpx list [/server] | List available tools | | mcpx schema /server <tool> | Show tool input schema | | mcpx /server <tool> --params '{}' | Execute a tool | | mcpx interactive [/server] | Start interactive REPL | | mcpx import [path] | Import from Claude Desktop config | | mcpx skills /server | Generate agent skill docs | | mcpx daemon start\|stop\|status | Manage connection daemon | | mcpx inspect /server | Show server capabilities and metadata | | mcpx prompts /server | List MCP prompt templates | | mcpx prompt /server <name> | Get a prompt template | | mcpx resources /server | List MCP resources | | mcpx resource /server <uri> | Read a resource | | mcpx diff /server | Compare tool schemas against snapshot | | mcpx test /server | Verify server health | | mcpx watch <interval> /server <tool> | Periodic re-execution | | mcpx workflow <file> | Run multi-step YAML workflow | | mcpx hook add\|list\|remove | Manage middleware hooks | | mcpx alias set\|list\|remove | Manage tool aliases | | mcpx run <name> | Execute a saved alias |

Invocation Patterns

# Slash-command (primary)
mcpx /server tool --params '{"key": "value"}'

# Per-field flags (human sugar)
mcpx /server tool --key value

# Single-shot from another tool
mcpx -p '/server tool --params \'{"key": "value"}\''

# Legacy exec mode
mcpx -s "npx @mcp/server" exec tool --key value

# Pipe output between tools
mcpx /pg execute_sql --params '{"sql":"SELECT id FROM users"}' | mcpx /pg get_column_cardinality --params-stdin

# Extract a specific field
mcpx /pg database_overview --field uptime

Output Formats

By default, all commands return a JSON envelope (agent-friendly). Use --format for human-readable output:

mcpx list /weather --format table    # tabular output
mcpx list /weather --format yaml     # YAML output
mcpx list /weather --format json     # JSON envelope (default)
mcpx list /weather --format csv      # CSV output
mcpx list /weather --format markdown # Markdown table

Output Contract

Agent-facing commands return a JSON envelope:

{"ok": true, "result": [{"type": "text", "text": "..."}]}
{"ok": true, "tools": [{"name": "...", "description": "...", "inputSchema": {...}}]}
{"ok": true}
{"ok": false, "error": {"code": 3, "message": "Missing required: city"}}

Success envelopes contain one of: result (tool output), tools (listing), schema (introspection), servers (registry), or no payload (add/remove/update confirmation).

Exit Codes

| Code | Meaning | |------|---------| | 0 | Success | | 1 | Tool execution error | | 2 | Connection error | | 3 | Validation error | | 4 | Config error | | 5 | Internal error |

Server Registration

# Register with env vars
mcpx add pg "npx @toolbox-sdk/server --prebuilt=postgres --stdio" \
  -e POSTGRES_HOST=localhost \
  -e POSTGRES_PASSWORD=secret

# Import from Claude Desktop
mcpx import

Authentication

mcpx supports multiple authentication methods for remote MCP servers:

Static Bearer Token

mcpx add api "https://mcp.example.com/v1" \
  -H "Authorization: Bearer sk-your-token"

Custom Headers

mcpx add api "https://mcp.example.com/v1" \
  -H "X-API-Key: your-key" \
  -H "X-Org-Id: org-123"

OAuth Client Credentials

mcpx add api "https://mcp.example.com/v1" \
  --oauth-client-id your-client-id \
  --oauth-client-secret your-secret \
  --oauth-scope "read write"

SSE Transport

mcpx add api "https://mcp.example.com/sse" --transport sse

Stdio with Environment Variables (existing)

mcpx add pg "npx @mcp/server-postgres" \
  -e POSTGRES_PASSWORD=secret

Connection Daemon

mcpx spawns a background daemon to cache MCP server connections:

mcpx daemon start    # Start daemon (auto-starts on first use)
mcpx daemon status   # Check if running
mcpx daemon stop     # Stop daemon

The daemon auto-exits after 5 minutes of inactivity.

MCP Protocol Features

mcpx exposes the full MCP protocol surface — not just tools:

mcpx inspect /pg              # server name, version, capabilities, instructions
mcpx prompts /pg              # list prompt templates
mcpx prompt /pg debug-query   # get a specific prompt
mcpx resources /pg            # list data resources
mcpx resource /pg <uri>       # read a resource

MCP Gateway

mcpx can act as a universal MCP aggregator, exposing all registered servers through a single connection:

mcpx serve                   # stdio (for Claude Desktop / Cursor)
mcpx serve --port 8080       # HTTP (for remote agents)
curl http://localhost:8080/health

Workflows

Run multi-step operations across servers with YAML workflow files:

# workflow.yaml
name: Daily Report
steps:
  - server: pg
    tool: execute_sql
    params: { sql: "SELECT count(*) as n FROM orders" }
    output: count
  - server: slack
    tool: send_message
    params: { text: "Orders today: {{count}}" }

mcpx workflow workflow.yaml

Steps execute sequentially. Use output to name variables, {{var}} to interpolate.

Hooks

Run shell commands before or after tool calls:

mcpx hook add 'before:pg.*' 'echo "$MCPX_TOOL" >> /var/log/mcpx.log'
mcpx hook add 'after:pg.execute_sql' 'notify-send "SQL executed"'
mcpx hook list
mcpx hook remove 'before:pg.*'

Hooks receive MCPX_SERVER, MCPX_TOOL, and MCPX_PARAMS environment variables.

Monitoring & Observability

mcpx watch 5s /pg list_active_queries       # poll every 5 seconds (NDJSON output)
mcpx test /pg                               # health check with latency
mcpx diff /pg                               # detect schema changes
mcpx --log /var/log/mcpx.ndjson /pg tool    # audit trail

Agent Skills

Generate SKILL.md documentation for any registered server:

mcpx skills /pg > SKILL-pg.md

Evaluation

A detailed comparison of direct MCP usage vs mcpx is available in docs/mcp-vs-mcpx-eval.md. It covers:

Automated test coverage (430 tests across 49 test files)
Latency benchmarks (direct MCP vs mcpx inline vs daemon-cached)
Token-budget estimates for agent consumption
A reproducible manual eval you can run end-to-end
Gap analysis and next steps

Development

npm run build        # Compile TypeScript
npm run dev          # Run via tsx
npm test             # Run all tests
npm run bench        # Benchmark direct MCP vs mcpx

License

MIT

MCP Servers