Skip to main content
skill CX Codex

mcp-cli · obra — Claude Code skill: on-demand MCP servers

A skill that has the agent call MCP servers through the mcp CLI, without wiring them in as permanent integrations.

Install into Claude Code · Skills CLI (Vercel) All methods and platforms ↓
npx skills add obra/superpowers-lab --skill mcp-cli -a claude-code
Open on GitHub

What mcp-cli does

The skill covers the mcp tool from mcptools. The agent inspects a server first, then makes the call it needs. Nothing gets registered as a standing integration, so your context window stays free of pre-loaded MCP definitions. Handy when you want to poke at an unfamiliar server before committing to it.

Everything runs in a fixed order. mcp tools <server-command> lists the tools, mcp resources and mcp prompts show data sources and prompt templates. Calls go through mcp call <tool_name> --params '<json>' <server-command>, and -f json or -f pretty changes the output. For repeat work there are aliases: mcp alias add fs npx -y @modelcontextprotocol/server-filesystem /home/user, stored in ~/.mcpt/aliases.json.

The binary has to live at ~/.local/bin/mcp; if it is missing, you build it from the mcptools source with Go. Make sure that directory is on your PATH before the first call. Servers that need auth take their token on the command line, like GITHUB_PERSONAL_ACCESS_TOKEN via -e, and Brave Search wants BRAVE_API_KEY. Transport is stdio, HTTP or SSE — SSE has to be named explicitly with --transport sse.

What it helps with

  • list the tools an unfamiliar MCP server exposes
  • make a one-off MCP call without editing any config
  • read a resource or prompt template from a server
  • debug an MCP server and check its parameter schema
  • set up an alias for a session and drop it afterwards

Example prompts for the agent

  • Show me the tools on npx -y @modelcontextprotocol/server-filesystem /tmp
  • Call read_file with {"path": "/tmp/example.txt"} on the filesystem server
  • Add an alias fs for the filesystem server and read README.md through it

How to install mcp-cli

1Skills CLI (Vercel)

npx skills add obra/superpowers-lab --skill mcp-cli -a claude-code

2Manually: put the files in place

~/.claude/skills/mcp-cli/SKILL.md

Copy the whole skill folder; for one project use .claude/skills/mcp-cli/

Step by step, with commands for each platform: How to install skills in Claude Code: five ways

Questions about mcp-cli

How do I install mcp-cli in Claude?

Run npx skills add obra/superpowers-lab --skill mcp-cli -a claude-code. The skill file lands at ~/.claude/skills/mcp-cli/SKILL.md.

What does mcp-cli do?

It teaches the agent to work with MCP servers through the mcp CLI: discover tools, resources and prompts, invoke them, read the results. No permanent integration required.

What do I need before using it?

The mcp binary at ~/.local/bin/mcp and that directory on your PATH. If the binary is missing, build it from the mcptools source with Go. Servers with auth need a token, such as GITHUB_PERSONAL_ACCESS_TOKEN.

Source text

SKILL.md · MIT
---
name: mcp-cli
description: Use MCP servers on-demand via the mcp CLI tool - discover tools, resources, and prompts without polluting context with pre-loaded MCP integrations
---

# MCP CLI: On-Demand MCP Server Usage

Use the `mcp` CLI tool to dynamically discover and invoke MCP server capabilities without pre-configuring them as permanent integrations.

## When to Use This Skill

Use this skill when you need to:
- Explore an MCP server's capabilities before deciding to use it
- Make one-off calls to an MCP server without permanent integration
- Access MCP functionality without polluting the context window
- Test or debug MCP servers
- Use MCP servers that aren't pre-configured

## Prerequisites

The `mcp` CLI must be installed at `~/.local/bin/mcp`. If not present:

```bash
# Clone and build
cd /tmp && git clone --depth 1 https://github.com/f/mcptools.git
cd mcptools && CGO_ENABLED=0 go build -o ~/.local/bin/mcp ./cmd/mcptools
```

Always ensure PATH includes the binary:
```bash
export PATH="$HOME/.local/bin:$PATH"
```

## Discovery Workflow

### Step 1: Discover Available Tools

```bash
mcp tools <server-command>
```

**Examples:**
```bash
# Filesystem server
mcp tools npx -y @modelcontextprotocol/server-filesystem /path/to/allow

# Memory/knowledge graph server
mcp tools npx -y @modelcontextprotocol/server-memory

# GitHub server (requires token)
mcp tools docker run -i --rm -e GITHUB_PERSONAL_ACCESS_TOKEN ghcr.io/github/github-mcp-server

# HTTP-based server
mcp tools https://example.com/mcp
```

### Step 2: Discover Resources (if supported)

```bash
mcp resources <server-command>
```

Resources are data sources the server exposes (files, database entries, etc.).

### Step 3: Discover Prompts (if supported)

```bash
mcp prompts <server-command>
```

Prompts are pre-defined prompt templates the server provides.

### Step 4: Get Detailed Info (JSON format)

```bash
# For full schema details including parameter types
mcp tools --format json <server-command>
mcp tools --format pretty <server-command>
```

## Making Tool Calls

### Basic Syntax

```bash
mcp call <tool_name> --params '<json>' <server-command>
```

### Examples

**Read a file:**
```bash
mcp call read_file --params '{"path": "/tmp/example.txt"}' \
  npx -y @modelcontextprotocol/server-filesystem /tmp
```

**Write a file:**
```bash
mcp call write_file --params '{"path": "/tmp/test.txt", "content": "Hello world"}' \
  npx -y @modelcontextprotocol/server-filesystem /tmp
```

**List directory:**
```bash
mcp call list_directory --params '{"path": "/tmp"}' \
  npx -y @modelcontextprotocol/server-filesystem /tmp
```

**Create entities (memory server):**
```bash
mcp call create_entities --params '{"entities": [{"name": "Project", "entityType": "Software", "observations": ["Uses TypeScript"]}]}' \
  npx -y @modelcontextprotocol/server-memory
```

**Search (memory server):**
```bash
mcp call search_nodes --params '{"query": "TypeScript"}' \
  npx -y @modelcontextprotocol/server-memory
```

### Complex Parameters

For nested objects and arrays, ensure valid JSON:

```bash
mcp call edit_file --params '{
  "path": "/tmp/file.txt",
  "edits": [
    {"oldText": "foo", "newText": "bar"},
    {"oldText": "baz", "newText": "qux"}
  ]
}' npx -y @modelcontextprotocol/server-filesystem /tmp
```

### Output Formats

```bash
# Table (default, human-readable)
mcp call <tool> --params '{}' <server>

# JSON (for parsing)
mcp call <tool> --params '{}' -f json <server>

# Pretty JSON (readable JSON)
mcp call <tool> --params '{}' -f pretty <server>
```

## Reading Resources

```bash
# List available resources
mcp resources <server-command>

# Read a specific resource
mcp read-resource <resource-uri> <server-command>

# Alternative syntax
mcp call resource:<resource-uri> <server-command>
```

## Using Prompts

```bash
# List available prompts
mcp prompts <server-command>

# Get a prompt (may require arguments)
mcp get-prompt <prompt-name> <server-command>

# With parameters
mcp get-prompt <prompt-name> --params '{"arg": "value"}' <server-command>
```

## Server Aliases (for repeated use)

If using a server frequently during a session:

```bash
# Create alias
mcp alias add fs npx -y @modelcontextprotocol/server-filesystem /home/user

# Use alias
mcp tools fs
mcp call read_file --params '{"path": "README.md"}' fs

# List aliases
mcp alias list

# Remove when done
mcp alias remove fs
```

Aliases are stored in `~/.mcpt/aliases.json`.

## Authentication

### HTTP Basic Auth
```bash
mcp tools --auth-user "username:password" https://api.example.com/mcp
```

### Bearer Token
```bash
mcp tools --auth-header "Bearer your-token-here" https://api.example.com/mcp
```

### Environment Variables (for Docker-based servers)
```bash
mcp tools docker run -i --rm \
  -e GITHUB_PERSONAL_ACCESS_TOKEN="$GITHUB_TOKEN" \
  ghcr.io/github/github-mcp-server
```

## Transport Types

### Stdio (default for npx/node commands)
```bash
mcp tools npx -y @modelcontextprotocol/server-filesystem /tmp
```

### HTTP (auto-detected for http/https URLs)
```bash
mcp tools https://example.com/mcp
```

### SSE (Server-Sent Events)
```bash
mcp tools http://localhost:3001/sse
# Or explicitly:
mcp tools --transport sse http://localhost:3001
```

## Common MCP Servers

### Filesystem
```bash
# Allow access to specific directory
mcp tools npx -y @modelcontextprotocol/server-filesystem /path/to/allow
```

### Memory (Knowledge Graph)
```bash
mcp tools npx -y @modelcontextprotocol/server-memory
```

### GitHub
```bash
export GITHUB_PERSONAL_ACCESS_TOKEN="your-token"
mcp tools docker run -i --rm -e GITHUB_PERSONAL_ACCESS_TOKEN ghcr.io/github/github-mcp-server
```

### Brave Search
```bash
export BRAVE_API_KEY="your-key"
mcp tools npx -y @anthropic/mcp-server-brave-search
```

### Puppeteer (Browser Automation)
```bash
mcp tools npx -y @anthropic/mcp-server-puppeteer
```

## Best Practices

### 1. Always Discover First
Before calling tools, run `mcp tools` to understand what's available and the exact parameter schema.

### 2. Use JSON Format for Parsing
When you need to process results programmatically:
```bash
mcp call <tool> --params '{}' -f json <server> | jq '.field'
```

### 3. Validate Parameters
The table output shows parameter signatures. Match them exactly:
- `param:str` = string
- `param:num` = number
- `param:bool` = boolean
- `param:str[]` = array of strings
- `[param:str]` = optional parameter

### 4. Handle Errors Gracefully
Tool calls may fail. Check exit codes and stderr:
```bash
if ! result=$(mcp call tool --params '{}' server 2>&1); then
  echo "Error: $result"
fi
```

### 5. Use Aliases for Multi-Step Operations
If making several calls to the same server:
```bash
mcp alias add tmp-server npx -y @modelcontextprotocol/server-filesystem /tmp
mcp call list_directory --params '{"path": "/tmp"}' tmp-server
mcp call read_file --params '{"path": "/tmp/file.txt"}' tmp-server
mcp alias remove tmp-server
```

### 6. Restrict Capabilities with Guard
For safety, limit what tools are accessible:
```bash
# Only allow read operations
mcp guard --allow 'tools:read_*,list_*' --deny 'tools:write_*,delete_*' \
  npx -y @modelcontextprotocol/server-filesystem /home
```

## Debugging

### View Server Logs
```bash
mcp tools --server-logs <server-command>
```

### Check Alias Configuration
```bash
cat ~/.mcpt/aliases.json
```

### Verbose Output
Use `--format pretty` for detailed JSON output to debug parameter issues.

## Quick Reference

| Action | Command |
|--------|---------|
| List tools | `mcp tools <server>` |
| List resources | `mcp resources <server>` |
| List prompts | `mcp prompts <server>` |
| Call tool | `mcp call <tool> --params '<json>' <server>` |
| Read resource | `mcp read-resource <uri> <server>` |
| Get prompt | `mcp get-prompt <name> <server>` |
| Add alias | `mcp alias add <name> <server-command>` |
| Remove alias | `mcp alias remove <name>` |
| JSON output | Add `-f json` or `-f pretty` |

## Example: Complete Workflow

```bash
# 1. Discover what's available
mcp tools npx -y @modelcontextprotocol/server-filesystem /home/user/project

# 2. Check for resources
mcp resources npx -y @modelcontextprotocol/server-filesystem /home/user/project

# 3. Create alias for convenience
mcp alias add proj npx -y @modelcontextprotocol/server-filesystem /home/user/project

# 4. Explore directory structure
mcp call directory_tree --params '{"path": "/home/user/project"}' proj

# 5. Read specific files
mcp call read_file --params '{"path": "/home/user/project/README.md"}' proj

# 6. Search for patterns
mcp call search_files --params '{"path": "/home/user/project", "pattern": "**/*.ts"}' proj

# 7. Clean up alias
mcp alias remove proj
```

## Troubleshooting

### "command not found: mcp"
Ensure PATH is set: `export PATH="$HOME/.local/bin:$PATH"`

### JSON parse errors
- Escape special characters properly
- Avoid shell expansion issues by using single quotes around JSON
- For complex JSON, write to a temp file and use `--params "$(cat params.json)"`

### Server timeout
Some servers take time to start. The mcp CLI waits for initialization automatically.

### Permission denied
For filesystem server, ensure the allowed directory path is correct and accessible.

© obra · MIT · obra/superpowers-lab

Similar skills

All skills for Agent setup: memory, context, orchestration →

mcp-cli by other authors comes first.

From the same set3

Project sponsors

Companies that keep this analytics open
Gold sponsor

Impressions, clicks and code copies — open statistics. We keep the numbers public for the same reason NeBlask itself is public. Become a sponsor