An AI mentor that keeps your project's specification up to date
- context
- EPITECH · team of 3
- period
- Sep → Nov 2026
- platforms
- Python · TypeScript · Ollama · Docker

> cat context.md
An EPITECH artificial intelligence project, built with Armand Lecomte and Rayan Ouerdane. My part: the CLI, the configuration, the agent profiles, the living specification, the prompt versions, the HTTP API, the web interface and the documentation. Armand built the document loaders and the RAG (still in progress), the code sandbox, the workspace files, the clarification questions and the benchmark. Rayan built the agent loop, the LLM client, the conversation service, the MCP tools and web search, and the vector database.
> cat how-it-works.md
- 01Wu, the default agent, guides you towards the answer. You can switch mid-conversation to Euclid (maths), Sensei (Japanese), Ada (code review) or Mira (product), or add your own profile in YAML.
- 02After each answer, a note-keeper returns only the sections of the specification that changed, as JSON edits the code applies to ASSISTANT.md, which is reloaded at the next start.
- 03In multi-agent mode, a coordinator hands parts of a task to specialised agents (research, code, files, review) described in a YAML file.
- 04The agents use tools: web search through MCP, the files of the workspace within set permissions, and a Docker sandbox to run and test code. When a missing choice blocks a request, the agent asks a closed question with options (human in the loop).
- 05The same agents run in the terminal or in a React web interface with streamed replies, through an HTTP API that also serves the profiles and the specification.
- 06Every interaction is traced locally (volume, latency, errors, tokens) and shown on a dashboard. wu-compare runs two prompt versions on the same messages, and a judge agent scores each reply.