Claude Code Codex Antigravity Cursor

driftguard

Your agent will build anything you ask. driftguard asks whether you should. It audits your AI-coding history, interviews you, and writes drift guards into your agent's rules.

Open the live demo GitHub ↗
Engraving of a seated figure lost in thought, surrounded by unfinished tools, with rays radiating outward
↑ Scroll up for the menu Ctrl K for commands
01 — What it does01 / 06

A bug fix turns into a new app. One side project becomes seven. Your agent never says no.

Agents
4Claude · Codex · Antigravity · Cursor
LLM tokens
0to scan, report, guard
Dependencies
0Python stdlib
Data
Localnothing uploaded
02 — How it works02 / 06

Four steps

03 — What it writes03 / 06

Rules your agent follows every session.

driftguard focus map: share of prompts that break focus, by prompt position, with an hour by weekday heatmap
01 Focus map

Analytics tools show you numbers. driftguard turns them into managed blocks in CLAUDE.md, AGENTS.md, GEMINI.md and .cursor/rules, so the next session stops you before the drift starts.

Drift guard

  • Project registry: active, stalled, parked, dead
  • Stop and poll when a task drifts
  • New-project gate: which project pauses?
  • Frustration protocol

Purpose blocks

  • Purpose and done-condition per project
  • The drift traps you fell into, with dates
  • Parked projects trigger the gate

Token crisis

  • Burn per 5-hour block
  • Frugal mode for subagents
  • A 40-line handoff, no LLM

Edits stay between driftguard's own markers · dry run by default · .bak backups

3D brain terrain: prompts per hour and weekday as towers, red where focus breaks
driftguard — FocusBrain terrain

Stop at prompt #2.
Not project #8.

Focus usually breaks right after the first answer. driftguard shows where yours does and writes the rule that catches it.

Tokens to scan
0
Handoff length
≤40
Agents
4
04 — Install04 / 06

One command. Then ask your agent.

The CLI installs with pipx, uv or pip. The one-liners also install the integrations for every agent they find on your machine.

  1. 01

    driftguard scan

    Read every history and pick a mode.

  2. 02

    /driftguard

    In your agent: scoring, interview, verdicts and guards.

  3. 03

    report --serve

    The interactive report on localhost.

  4. 04

    handoff

    Out of tokens? A 40-line HANDOFF.md, then a fresh session.

05 — Demo05 / 06

The report, on synthetic data.

Motivation timeline, prompt coaching, focus map, tokens, a project kill or keep board, a drift graph and 3D views. One HTML file that opens offline.

Open the live demo ↗
3D brain terrain view of the demo report
01 3D views
Focus map view of the demo report
02 Focus map
06 — Use it with the others06 / 06

Run it

For token numbers, use ccusage. For a friction report, use /insights. For an agent that stops you starting project #8, use driftguard.