A CLI client that argues MCP belongs in the shell

What it does

mcpc is a command-line client for the Model Context Protocol. It translates MCP operations—tools, resources, prompts, async tasks, and authentication—into ordinary shell commands. Humans can use it to inspect and debug servers interactively, while AI agents can treat it as a universal adapter that exposes any MCP server through standard shell primitives.

The interesting bit

The project’s core argument is that the CLI is the ideal local boundary between agents and MCP. Rather than injecting tool definitions into context windows, an agent only needs one Bash() capability; mcpc handles sessions, OAuth 2.1 flows, and even experimental x402 payments on the remote side. It also proxies connections to keep credentials away from AI-generated code.

Key highlights

• Maintains multiple persistent, stateful sessions to different servers simultaneously.
• Outputs JSON for composition with jq and shell pipelines in “code mode.”
• Discovers and connects servers from standard config files (Claude Desktop, Cursor, Windsurf, etc.) in bulk.
• Stores credentials in the OS keychain, falling back to a file-based store on headless systems.
• Includes an experimental proxy mode to sandbox MCP connections from untrusted agent code.

Caveats

• x402 payment support is currently experimental.
• Bulk auto-discovery skips local stdio servers unless explicitly enabled.
• Argument parsing for tool calls is strict: no spaces around the := separator.

Verdict

Developers building AI agents or orchestrating MCP servers from scripts should keep this nearby; if you prefer thick GUI clients or your agent doesn’t have shell access, it won’t help.