Each of the 8 tools takes a single query and returns a cognitive operation — abstract procedure steps and a reasoning topology DAG — which the model is then expected to run itself. The adaptive variants (adaptive-reasoning, adaptive-code, adaptive-anti-deception and adaptive-memory) put an adapter LLM in front of that, concretising the same steps into the vocabulary of the task at hand; adaptive-anti-deception draws from a library of 139 operations across six sub-layers. Because what comes back is a method rather than a result, the quality of the final reply still rests on the model that follows it, and every call is a round trip to the service.
An MCP server that returns cognitive operations for a model to run before it replies, rather than returning an answer itself.
- A reasoning operation for diagnostic, planning, trade-off and architecture questions, fetched before the answer is written.
- A code operation for generating, refactoring, reviewing or debugging, including prompts that carry a code block to act on.
- An anti-deception operation for requests that press for agreement, manufacture urgency, appeal to authority, or ask for something to be certified without evidence.
- A memory operation that sharpens an observation already formed about conversation state, user behaviour, drift or cross-turn patterns.
- An adaptive counterpart to each of those four, where an adapter LLM rewrites the procedure steps and topology nodes in the language of the specific task.
EJENTUM_API_KEY. Every tool sends its query to the service that returns the operation, and the adaptive variants put an adapter LLM in that path as well.
One command plus a key — npm install ejentum-mcp, then supply credentials