Comparison
How mcpx compares to other approaches for giving AI agents access to MCP tools.
MCPX vs Native MCP
| Aspect | Native MCP | MCPX |
|---|---|---|
| Context cost | 10-30K tokens per server, upfront | 0 tokens upfront |
| Discovery | All schemas loaded at session start | On-demand via mcpx list |
| Startup | Every session pays init cost | Daemon mode: init once |
| Max servers | ~5-10 before context fills | Unlimited |
| Composability | Server-to-server: not possible | Pipe stdout between servers |
| Integration | Client-specific (Claude Code, Cursor) | Any tool that runs Bash |
Choose native MCP when:
- You use 1-2 small servers with few tools
- You need bidirectional communication (server-initiated notifications)
- Your AI client doesn't support Bash execution
Choose mcpx when:
- You use 3+ servers or servers with many tools
- Context efficiency matters
- You want tools to work across AI clients
- You need server-to-server composition
MCPX vs REST API Wrappers
Some people wrap MCP servers behind REST APIs and call them via curl.
| Aspect | REST Wrapper | MCPX |
|---|---|---|
| Setup | Write HTTP server + deploy | mcpx init |
| Discovery | Custom (Swagger/OpenAPI) | Built-in (--help, list) |
| Auth | You implement it | OS keychain |
| Process management | You manage it (systemd, Docker) | Daemon mode built in |
| Schema mapping | Manual | Automatic from MCP |
mcpx is the REST wrapper approach without the REST wrapper.
MCPX vs Direct Subprocess Calls
You could have the AI spawn MCP servers directly via exec.Command equivalent:
| Aspect | Direct Subprocess | MCPX |
|---|---|---|
| Handshake | AI must implement MCP init | mcpx handles it |
| Schema parsing | AI parses JSON schema | Flags auto-generated |
| Connection reuse | New process every call | Daemon mode |
| Error handling | AI parses JSON-RPC errors | Exit codes + stderr |
| Config | Hardcoded in AI instructions | YAML config files |
Context Cost Breakdown
Real-world measurements of schema sizes:
| Server | Tools | Schema Size | Native MCP Cost |
|---|---|---|---|
| Serena | 21 | ~18K chars | ~20K tokens |
| Sequential Thinking | 1 | ~2K chars | ~3K tokens |
| Filesystem | 11 | ~8K chars | ~10K tokens |
| Brave Search | 2 | ~3K chars | ~4K tokens |
| GitHub | 30+ | ~25K chars | ~30K tokens |
With mcpx, all of these cost 0 tokens until the AI actually needs them. A single mcpx list serena -v call costs ~500 tokens in the response — and only when needed.