Skip to content

Getting Started

Sensez is meant to run inside an agent loop so it can catch smells before they reach a pull request. The fastest path is to keep it close to the edit: use it inside your coding agent loop to provide it quick feedback on its work, as it's getting developed.

1. Install

# As a project dev dependency
uv add --dev sensez

# Or as a global CLI
uv tool install sensez

npm install --save-dev sensez

2. Initialize the repository

uv run sensez init

npx sensez init

sensez init setup guides you through the install, after you select which coding agent to use it with. Then:

  • writes a commented starter config — sensez.toml, or a [tool.sensez] section in pyproject.toml if you prefer;
  • registers the Sensez MCP server with your agent, so the agent launches Sensez automatically at startup;
  • installs the Sensez skill that teaches the agent when and how to scan;
  • creates .sensez/ for local metrics and caches, and adds it to .gitignore, if local metrics are enabled;
  • optionally installs an experimental end-of-turn gate hook (Claude Code only).

What init writes for each supported agent:

Agent MCP config Agent skill Gate hook
Claude Code .mcp.json .claude/skills/sensez Optional, in .claude/settings.json
Cursor .cursor/mcp.json — —
Cline .cline/mcp.json .cline/skills/sensez —
Codex .codex/config.toml .codex/skills/sensez —
OpenCode opencode.jsonc .opencode/skills/sensez —
Pi .pi/mcp.json .pi/skills/sensez —
Other / none — (init prints guidance instead) — —

3. Restart your agent

Agents load MCP servers at startup. Restart the agent (or reload the window) after init, or the sensez tools will not appear.

4. Tune it

Everything Sensez does is configured in the file init wrote. Continue with the Configuration Reference for action levels, per-language overrides, boundary rules, and accepting findings, and the Finding Reference for what each detector does.

Next