MCP Client

MCP Client (Universal)

This skill provides a small, portable CLI for connecting to MCP (Model Context Protocol) servers on-demand:

• list configured servers
• list tools for a single server (progressive disclosure)
• call a tool with JSON arguments

It is designed to avoid pasting large tool schemas into the agent context up front.

Where It Lives

• When installed into an agent home (recommended): $ICA_HOME/skills/mcp-client/ (or ~/.codex/skills/mcp-client/, ~/.claude/skills/mcp-client/)
• From this repo (development/reference): src/skills/mcp-client/

Recommendation: Use mcp-proxy For “Register Once”

If you want to register only a single MCP server in your agent runtime and manage upstream servers + auth centrally, use the mcp-proxy skill. It mirrors upstream tools as <server>.<tool> and provides stable broker tools under proxy.*.

Script:

python "$ICA_HOME/skills/mcp-client/scripts/mcp_client.py" --help

Dependencies

You need Python 3.9+ and:

pip install mcp

Configuration

The client resolves config in this order:

1. MCP_CONFIG_PATH env var (path to JSON)
2. MCP_CONFIG env var (inline JSON)
3. Merge project + ICA home (default):
   • project: .mcp.json
   • user: $ICA_HOME/mcp-servers.json (or $ICA_HOME/mcp.json)
   • default precedence: project overrides user (flip with ICA_MCP_CONFIG_PREFER_HOME=1)
4. ~/.claude.json (Claude Code compatibility fallback; reads mcpServers if no other config exists)

To get started, copy:

• references/example-mcp-config.json -> .mcp.json (project-local), OR
• references/example-mcp-config.json -> $ICA_HOME/mcp-servers.json (agent-home local)

Security note:

• Do not commit real API keys. Keep .mcp.json and references/mcp-config.json local-only.
• If you need an API key and the user hasn't provided it yet, ask for it.

Authentication (Typical Patterns)

Most MCP servers require authentication. This client does not run OAuth flows for you; it only sends what you configure.

Common patterns:

• Remote HTTP MCP servers: set headers.Authorization to Bearer <token> (or use another header like X-API-Key).
• Convenience: api_key is treated as sugar for headers.Authorization = "Bearer <api_key>".
• Local stdio MCP servers: set required secrets in env (for example GITHUB_TOKEN, POSTGRES_CONNECTION_STRING).
• Environment placeholders: strings containing ${VAR} are expanded from your process environment at runtime (best-effort).

OAuth (Interactive)

If a server uses OAuth (not static API keys), configure an oauth block for that server, then run:

python "$ICA_HOME/skills/mcp-client/scripts/mcp_client.py" auth <server_name>

Supported OAuth styles:

• Authorization Code + PKCE (browser-based redirect to localhost)
• Device Code (copy/paste a code into a browser)

Tokens are stored locally in $ICA_HOME/mcp-tokens.json and will be used automatically for calls if the server does not already specify headers.Authorization.

OAuth config (PKCE, explicit endpoints):

{
  "mcpServers": {
    "remote-oauth": {
      "url": "https://example.com/mcp",
      "oauth": {
        "type": "pkce",
        "authorization_url": "https://idp.example.com/oauth/authorize",
        "token_url": "https://idp.example.com/oauth/token",
        "client_id": "YOUR_CLIENT_ID",
        "scopes": ["openid", "profile", "offline_access"],
        "redirect_uri": "http://127.0.0.1:8765/callback"
      }
    }
  }
}

OAuth config (OIDC discovery + PKCE):

{
  "mcpServers": {
    "remote-oidc": {
      "url": "https://example.com/mcp",
      "oauth": {
        "type": "oidc_pkce",
        "issuer": "https://idp.example.com/",
        "client_id": "YOUR_CLIENT_ID",
        "scopes": ["openid", "profile", "offline_access"]
      }
    }
  }
}

OAuth config (device code):

{
  "mcpServers": {
    "remote-device": {
      "url": "https://example.com/mcp",
      "oauth": {
        "type": "device_code",
        "device_authorization_url": "https://idp.example.com/oauth/device/code",
        "token_url": "https://idp.example.com/oauth/token",
        "client_id": "YOUR_CLIENT_ID",
        "scopes": ["read", "write"]
      }
    }
  }
}

Commands

# List configured servers
python "$ICA_HOME/skills/mcp-client/scripts/mcp_client.py" servers

# List tools (with full parameter schemas) from one server
python "$ICA_HOME/skills/mcp-client/scripts/mcp_client.py" tools <server_name>

# Call a tool
python "$ICA_HOME/skills/mcp-client/scripts/mcp_client.py" call <server_name> <tool_name> '{"arg":"value"}'

Example: Remote MCP (Bearer Auth)

python "$ICA_HOME/skills/mcp-client/scripts/mcp_client.py" servers
python "$ICA_HOME/skills/mcp-client/scripts/mcp_client.py" tools <remote-server>
python "$ICA_HOME/skills/mcp-client/scripts/mcp_client.py" call <remote-server> <tool_name> '{"param":"value"}'

Example: Sequential Thinking (Local stdio MCP server)

python "$ICA_HOME/skills/mcp-client/scripts/mcp_client.py" tools sequential-thinking
python "$ICA_HOME/skills/mcp-client/scripts/mcp_client.py" call sequential-thinking sequentialthinking '{"thought":"Breaking down the problem...","thoughtNumber":1,"totalThoughts":5,"nextThoughtNeeded":true}'

Config Format

The config file can be either:

• {"mcpServers": { ... }} (recommended), or
• { ... } where the top-level keys are server names.

{
  "mcpServers": {
    "remote-example": {
      "url": "https://example.com/mcp",
      "headers": {
        "Authorization": "Bearer ${REMOTE_API_KEY}"
      }
    },
    "sequential-thinking": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-sequential-thinking"]
    }
  }
}

Transport detection:

• command (+ optional args) -> stdio
• url ending in /sse -> sse
• url ending in /mcp -> streamable_http
• api_key -> sugar for headers.Authorization = "Bearer <api_key>"

References

• references/mcp-servers.md - Common server configurations
• references/example-mcp-config.json - Template config file
• references/python-mcp-sdk.md - Python SDK notes and examples