The trick is a clean split of duties: the service manager owns the weights, the client owns the history. That is what lets you plan with a reasoning model and then implement with a coding model in the same thread. Set `health_timeout` above its default of 30 seconds if your models are large — a swap that times out is almost always weights still loading.
A model switcher for a local llama-server. It starts and stops model services through launchctl on macOS or systemd on Linux, so the weights behind the endpoint change while the client keeps the full message history it re-sends every turn.
- `list_models` lists every configured model with its load status and the current mode
- `get_current_model` returns the alias of the model loaded right now
- `swap_model` unloads the current model, loads the one you name, and waits for the health check to pass
- `create_model_config` writes a new launchd plist or systemd unit for a model you describe in the chat
- Two resources, `llama-swap://config` and `llama-swap://status`, plus a `swap-workflow` prompt for the plan-then-implement pattern
The PyPI package mcp-llama-swap, version 1.0.0, run as `uvx mcp-llama-swap`. Python 3.10+, llama-server on the machine, and launchctl or systemd. Point LLAMA_SWAP_CONFIG at an absolute path to your config.json, which maps aliases to service filenames. Claude Code has to be pointed at a LiteLLM proxy for the swap to be invisible to it.
One command — uvx mcp-llama-swap