Lohnt sich dein eigenes Cursor?

Build me an AI coding setup to replace Cursor, assembled from parts I own.
Requirements:

- Do not write an editor or an agent. Install and wire together: VS Code or
  VSCodium, the Continue extension for in-editor chat, edit, and
  autocomplete, and a terminal agent (aider or Claude Code) for multi-file
  jobs.
- Continue config at ~/.continue/config.yaml: chat and edit on Claude
  Sonnet, autocomplete on a cheap fast model (Codestral, or qwen2.5-coder
  via Ollama to keep completions local and free).
- All API keys read from environment variables set in my shell profile,
  nothing hardcoded in any config file.
- Enable Continue's codebase indexing on my main repo and verify a
  where-is-this-defined question retrieves the right file.
- Keybindings mapped to muscle memory: cmd+K edit selection, cmd+L chat,
  Tab accepts completions.
- A spend check in the README: per-token API pricing vs the $20 flat fee,
  and a reminder to compare the provider dashboards after two weeks of real
  use.
- Out of scope: cloud and background agents, Cursor's composer UX, and team
  features; that iteration speed is exactly what the subscription buys.
- README: install commands per OS, the full config file, and how to swap
  models later.