Tenchi-MCP

The models used can be configured via a TOML file. When doing so, make sure the model names match those that appear in `ollama list`. Role and parameter settings are also configurable.

Since it's an MCP server, theoretically, if it fits in memory, all operations can be performed with local LLMs.

The workflow is as follows:

1. The cloud LLM considers which model and prompt to use, then makes a call
2. The implementation role writes out the code
3. The review role checks it
4. Return to the cloud LLM, and if it's usable, use it as-is; if there are issues, make corrections (sometimes returning to step 2)

Even just creating this kind of flow should change token efficiency.

Generating from 0 to 1 consumes quite a lot of tokens, but by pre-generating from 0 to 0.5 and having the cloud LLM side focus on corrections and refinements, we can improve token efficiency somewhat. However, there is the drawback that inference can take a very long time depending on the environment, so careful judgment about when to use it is necessary.

※The name comes from cloud → sky (空) and local → region (地), creating "Tenchi MCP" (天地MCP - Heaven and Earth MCP).