CodeSergeantYour AI drill sergeant for deep-work sessions. It reads what app and window you're actually in, compares it to the objective you set, and cuts in with a voice line before drift eats the block. Local-first judgment on Apple Silicon.
1// FocusDaemon.swift 2import SwiftUI 3import Combine 4 5final class FocusDaemon: ObservableObject { 6 @Published var objective: String 7 @Published var state: Session = .active 8 9 func tick() async { 10 let v = await sergeant.judge() 11 if v.isDrift { 12 voice.speak("Eyes on it.") 13 } 14 } 15}
Session active
Code Sergeant works when it feels watchful, opinionated, and slightly confrontational. Three behaviors drive the entire product.
Your objective gets compared against the live app + window title every few seconds. A local LLM scores whether you're on-mission, so “YouTube · Python Tutorial” stays quiet and “YouTube · Cat Videos” gets called out.
OBJ-LOCKLaunch a session from the menu bar, keep the app out of the way, and let it watch your active context while you work.
MENU-BARVoice feedback, reminders, and personalities give the product presence, which is the whole point when focus slips.
VOICE-01Three views from the macOS app. Mission setup, active surveillance, and the command-tuning panel.
Set The Objective
Set the objective, lock in the work block, and make the session explicit before you start. No vague timers. A stated mission the Sergeant can hold you to.
During The Session
Every ~10 seconds, CodeSergeant asks a local LLM whether your current window is on-mission. A 30-second cooldown prevents nagging, and sustained off-task patterns escalate into a voice warning. On hard drift, a strobe overlay.
Tune The Sergeant
Pick the Ollama model, pick the personality, tune the Pomodoro and XP loop. Voice can stay on-device via macOS TTS, or you can wire an ElevenLabs key for a richer sergeant.
Sessions run on configurable Pomodoro cycles. Completing a block earns XP against a rank ladder. Voice features stay out of the way until you invoke them.
Configurable work and break durations. The Sergeant announces transitions so you never miss the start of a block.
POMODOROEvery completed focus block earns XP. Ranks climb with accumulated sessions. Visible, animated, and persistent.
XP-RANK“Hey Sergeant” to chat during a session. “Take note Sergeant” to capture a transcribed note without touching the keyboard. Music and browser audio duck while the Sergeant speaks.
WAKE-WORDTwo TTS paths: Apple's on-device voices for a fully local setup, or an ElevenLabs API key for studio-grade delivery. Every line is written to read like an NCO who's watched you blow the last three sessions, not a polite assistant.
voice-test://sergeant · channel 01
"Eyes on the objective, soldier. You said deep work. Act like it."
The useful feedback comes from actually running a session that matters, not clicking around for five minutes.
Pick a task you actually need to finish and let the app watch what happens when your attention starts to bend.
Try the voice, settings, and nudges hard enough to learn whether the tone motivates you or becomes noise.
The useful feedback is where the app judged context wrong, interrupted at the wrong moment, or failed to apply pressure.
By default, context judgment runs on-device via Ollama and voice uses macOS's built-in TTS, with no cloud round-trip. If you add an ElevenLabs or OpenAI API key in Settings, those providers are called directly from your machine for voice and/or judgment. Nothing else is transmitted: no telemetry, no analytics pings.
Every ~10 seconds it reads your frontmost app and window title via macOS Accessibility APIs, and asks a local LLM whether that matches your stated objective. A 30-second cooldown prevents nagging, and a rolling pattern detector escalates when drift sticks. Research-shaped work near the mission stays quiet.
macOS 14 or newer on Apple Silicon (M1 or later). Ollama is optional but recommended for AI judgment. Without it, a rule-based fallback runs. ElevenLabs and OpenAI are optional and key-based. Intel Macs are not supported.
Yes. Four stock personalities ship: Sergeant, Buddy, Advisor, and Coach. Each has its own wake word and phrase library. You can also define a Custom personality (name, wake word, tone) and paste an ElevenLabs voice ID to drive the sound.
No. Timers don't judge what you're doing. Code Sergeant watches context against a stated mission and applies pressure when drift is real. A timer is a stopwatch; this is a drill sergeant.
Download the beta, read the setup guide, and send feedback after you've tried to break your own focus with it.
