Native tools vs MCP
When to expose actions as SDK tools versus running the MCP server.
Nolixan exposes the same governed actions two ways. Both apply policy and audit; the choice is about how your runtime calls them.
Native SDK tools
The agent-framework adapters hand actions to your LLM as native tools (OpenAI / Anthropic / Gemini / LangChain / CrewAI / Vercel AI) and run the model's tool-calls back through the action API.
Use when you control the agent loop and call an LLM API directly — you want tools in your code, in your language, with no extra process.
MCP server
The MCP server exposes actions over JSON-RPC to any MCP client.
Use when the runtime is an MCP host you don't control the loop of — Claude Desktop, an IDE, or a third-party agent — and you just need to point it at an endpoint.
Rule of thumb
| Native tools | MCP | |
|---|---|---|
| You own the agent loop | ✓ | — |
| Host is an existing MCP client | — | ✓ |
| Language-native, no extra process | ✓ | — |
| Same governance (policy + audit) | ✓ | ✓ |
Tool naming provider__action | ✓ | ✓ |