Skip to content

Comparison

How mcpx compares to other approaches for giving AI agents access to MCP tools.

MCPX vs Native MCP

AspectNative MCPMCPX
Context cost10-30K tokens per server, upfront0 tokens upfront
DiscoveryAll schemas loaded at session startOn-demand via mcpx list
StartupEvery session pays init costDaemon mode: init once
Max servers~5-10 before context fillsUnlimited
ComposabilityServer-to-server: not possiblePipe stdout between servers
IntegrationClient-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.

AspectREST WrapperMCPX
SetupWrite HTTP server + deploymcpx init
DiscoveryCustom (Swagger/OpenAPI)Built-in (--help, list)
AuthYou implement itOS keychain
Process managementYou manage it (systemd, Docker)Daemon mode built in
Schema mappingManualAutomatic 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:

AspectDirect SubprocessMCPX
HandshakeAI must implement MCP initmcpx handles it
Schema parsingAI parses JSON schemaFlags auto-generated
Connection reuseNew process every callDaemon mode
Error handlingAI parses JSON-RPC errorsExit codes + stderr
ConfigHardcoded in AI instructionsYAML config files

Context Cost Breakdown

Real-world measurements of schema sizes:

ServerToolsSchema SizeNative MCP Cost
Serena21~18K chars~20K tokens
Sequential Thinking1~2K chars~3K tokens
Filesystem11~8K chars~10K tokens
Brave Search2~3K chars~4K tokens
GitHub30+~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.

Released under the MIT License.