The idea is sound and unusual: a research task that would burn thousands of tokens of tool chatter runs in another process and returns a paragraph. The cost is honest in the README, which opens by saying it takes some setting up — you are configuring and maintaining a whole second MCP client before the first call works. The parallel tool is also flagged as waiting on an upstream `mcp-client-cli` change, so exercise it on a short list before building anything on top of it.
An MCP server that is also an MCP client. It sends a question to a separate LLM running its own tool stack, discards every intermediate step that model took, and returns just the result — so the work happens somewhere else and your context window only receives the answer.
- A question answered by a separate LLM, with all of its intermediate tool calls dropped and only the final output returned — `execute_mcp_client`
- One prompt run across a list of inputs at once, such as asking the same question of six cities — `execute_parallel_mcp_client`
- A list processed in parallel and then folded to a single result, with a `mapPrompt` carrying an `{item}` placeholder and a `reducePrompt` carrying `{accumulator}` and `{result}`, plus an optional starting value — `execute_map_reduce_mcp_client`
The delegated model is your own: this drives `mcp-client-cli`, which you install separately along with its config at `~/.llm/config.json` and whichever MCP servers that client should have. You also write a small shell script that activates its virtualenv and runs the `llm` executable, then point `MCP_INCEPTION_EXECUTABLE` at that script and `MCP_INCEPTION_WORKING_DIR` at the client's working directory. The npm package name is `mcp-inception` (0.1.0) and it is marked private, so build from a clone with `npm install` and `npm run build`.
One command — npx -y github:tanevanwifferen/mcp-inception