Skip to main content
MCP is a well-engineered protocol. It has a clear spec, good SDKs, and broad adoption. So why are people looking for alternatives?

​
Complexity cost

To expose a simple function via MCP, you need to:
  1. Pick an SDK (TypeScript, Python, etc.)
  2. Implement a server with JSON-RPC handlers
  3. Define tool schemas with JSON Schema
  4. Handle lifecycle (initialize, capabilities, shutdown)
  5. Configure transport (stdio or HTTP)
  6. Run and manage the server process
To expose the same function via SCP:
  1. Write a SKILL.md describing what it does
  2. Write a script that does it

​
The server problem

MCP requires a running server for every integration. Each server is a process to manage, monitor, and restart. For an agent with 10 integrations, that’s 10 server processes. SCP skills are inert files until the agent decides to use them. Nothing runs until something needs to run. When it does run, it’s a short-lived process that exits when done.

​
LLMs read markdown better than schemas

MCP describes tools with JSON Schema. SCP describes skills in markdown. Modern LLMs are remarkably good at understanding natural language instructions — often better than they are at correctly interpreting formal schemas. A well-written SKILL.md gives the agent context that a JSON schema can’t: when to use the skill, why it exists, caveats to watch for, and examples of correct usage.

​
What MCP gets right

Not everything about MCP is wrong. Some things are genuinely useful:
  • Standardized tool descriptions — so any client can discover capabilities
  • Structured input/output — so tools have predictable interfaces
  • Broad adoption — so there’s a large ecosystem
SCP achieves the first two through simpler means (markdown for descriptions, stdin/stdout for I/O). The ecosystem is what we’re building now.

​
When MCP makes sense

MCP is a better fit when you need:
  • Remote tool servers serving many clients over the network
  • Fine-grained capability negotiation between client and server
  • Stateful, long-running connections with bidirectional communication
If your use case is “let an AI agent on my machine use a tool,” SCP is simpler.

​
Side by side

SCPMCP
Time to first skill2 minutes30+ minutes
Runtime dependenciesNoneSDK + runtime
Running processes0 (until invoked)1 per integration
Lines of code~10 (SKILL.md + script)~100+ (server + handlers)
Learning curveWrite markdown, write scriptsLearn protocol, SDK, transports
DiscoveryRead a fileJSON-RPC tools/list
DebuggingRun the script manuallyInspect JSON-RPC messages