projects
software · now
nemonik
Give coding agents a repeatable workflow, the right tools for each step and the context to continue where the last session left off.
The problem
My agent had my workflow and still skipped steps. The more rules and context I added, the less consistently it followed them. The full story is in my first post.
How it's built
- CLI · Rust
- nk run starts Claude Code or Codex in its own tmux session with a kit's system prompt, skills, MCP servers, hooks and workflow in place. Skills go into a folder for that session, outside your repo. The folder is removed when the session ends. It ships as one signed binary through Homebrew or an install script.
- MCP server · Python
- Holds the context, memory, journal and status for each project, plus kits, prompts and skills. Search combines full-text matching with text embeddings, so agents find things by meaning. Claude, ChatGPT and Codex connect to it over OAuth.
- API · Scala
- Owns the database schema and the login flow (OAuth 2.1 with PKCE). Postgres row-level security decides who can read or edit each kit, prompt, skill and project.
- Web app · React, TypeScript
- Where kits, prompts, skills and workflows are written and edited. A project's context can be read and searched there too.
- Infrastructure · Postgres with pgvector, Redis, Kubernetes
- Postgres stores everything, including the embeddings. It runs on Kubernetes, deployed through ArgoCD.