Explored AI Miles project overview and assessment
Explored AI Miles project overview and assessment
Agent
Cursor
Model
Composer 2.5
Started
Jun 19, 8:32 PM
Messages45
Downloads0
Summarized AI Miles project purpose.
UnderstandOutlined strengths of AI Miles.
UnderstandReviewed and explained weaknesses of AI Miles.
Understandhi, can u tell me what this is all about?
Exploring the workspace to summarize what this project is about.
{
"success": true,
"finalTodos": [
{
"content": "Explore project structure and key files",
"status": "in_progress",
"id": "1"
},
{
"content": "Summarize what the project is about for the user",
"status": "pending",
"id": "2"
}
]
}{
"success": true,
"finalTodos": [
{
"content": "Explore project structure and key files",
"status": "completed",
"id": "1"
},
{
"content": "Summarize what the project is about for the user",
"status": "in_progress",
"id": "2"
}
],
"initialTodos": [
{
"content": "Explore project structure and key files",
"status": "in_progress",
"id": "1"
},
{
"content": "Summarize what the project is about for the user",
"status": "pending",
"id": "2"
}
],
"wasMerge": true
}{
"success": true,
"finalTodos": [
{
"content": "Explore project structure and key files",
"status": "completed",
"id": "1"
},
{
"content": "Summarize what the project is about for the user",
"status": "completed",
"id": "2"
}
],
"initialTodos": [
{
"content": "Explore project structure and key files",
"status": "completed",
"id": "1"
},
{
"content": "Summarize what the project is about for the user",
"status": "in_progress",
"id": "2"
}
],
"wasMerge": true
}AI Miles is a local macOS app that tracks how much you use AI coding tools — then turns that into gamified stats, sponsor placements, and (eventually) social ranking and rewards.
Think of it as Strava for AI-assisted coding: tokens become “miles,” you get streaks, levels, achievements, and a monthly “Wrapped” report — all computed from your machine, without reading your prompts or code.
Tracks usage locally — It scans metadata from tools like Claude Code, Codex, Gemini, Aider, and Copilot CLI (timestamps, model names, token counts). Everything stays in ~/.aimiles/aimiles.db. No prompt/response content is sent anywhere.
Gamifies your AI usage — Token usage is converted into miles, with streaks, tiers, heatmaps, quests, and a monthly Wrapped report (similar to Spotify Wrapped).
Shows sponsor ads in “dead air” — While AI tools are thinking/spinning, AI Miles can show rotating sponsor messages (Ramp, Vanta, Linear, etc.) via:
aimiles run -- codex) that rewrites spinner outputRewards catalog — rewards.json defines perks you could redeem with miles (credits, free months, swag), though redemption likely needs the backend.
| Piece | Purpose |
|---|---|
aimiles/ | Core Python package — ingest logs, compute stats, CLI, dashboard |
install.sh | One-command setup: shell aliases, VS Code extension, menu bar |
vscode-extension/ | Status-bar widget for editors |
browser-extension/ | Browser-side piece (limited coverage) |
BACKEND.md | Spec for a future multiplayer layer — leaderboards, friends, teams |
COVERAGE.md | Which tools/environments can be reached and how |
Right now it’s single-player (all local). The plan in BACKEND.md is to add a server for global rank, friends, activity feed, and teams (“Team Claude vs Team GPT”) — syncing only daily aggregates, not raw logs.
In one line: AI Miles monetizes and gamifies AI coding usage by turning local token stats into miles, streaks, and sponsor visibility — with a path toward social leaderboards and redeemable rewards.
If you want, I can walk through how to run it, how the scoring works, or how a specific piece (ingest, ads, Wrapped) is implemented.
what are its strengths
Reviewing a few more parts of the codebase so I can give a grounded take on its strengths.
I am synthesizing the strengths from both product and technical perspectives.
Privacy-first architecture with local databases and explicit metadata-only handling builds strong trust. Multi-tool aggregation across fragmented AI platforms solves a real pain point.
AI Miles has several real strengths — both as a product idea and in how this repo is built.
It only ingests metadata (timestamps, models, token counts) from logs tools already write locally. It deliberately avoids prompt/response content, stores everything in a local SQLite DB, and the planned backend syncs daily aggregates only. That’s a strong trust story in a space where people are wary of yet another tracker.
Developers use many AI tools at once — Claude Code, Codex, Gemini, Aider, Copilot, Continue, Cursor, etc. AI Miles pulls them into one unified score (miles, streaks, heatmap, Wrapped). That’s a genuine pain point nobody else really owns yet.
Most products can reach one environment. AI Miles has thought through the full stack:
aimiles run -- codex)COVERAGE.md is unusually honest about what works, what’s fragile, and what’s impossible. That clarity is a strength in itself.
The gamification isn’t random — it mirrors what works elsewhere:
“Rewards get the install. Status creates the addiction.” is a sharp product insight: perks drive adoption; rank and social comparison drive retention.
Three layers that reinforce each other:
Each layer has a concrete implementation already started locally.
Works without an account today. install.sh is one command, idempotent, and reversible. Ingest is incremental (byte offsets, deduped events) so re-runs stay cheap. That’s important for something that needs to run quietly in the background.
Miles aren’t naive token counts. Output and reasoning tokens are weighted much higher than cache reads, so inflating your score by loading huge cached contexts doesn’t work. Miles are computed at query time from raw counts, so scoring can be tuned without re-ingesting history.
The binary patcher and PTY proxy could easily become dangerous footguns. Instead they have:
That’s mature engineering for a concept that could otherwise feel reckless.
BACKEND.md defines a clean sync contract that matches what the CLI already emits (aimiles --json, daily rollups). The local app is fully usable solo; multiplayer is an additive layer, not a rewrite.
Bottom line: Its biggest strength is the combination of trust (local/metadata-only), breadth (many tools + many surfaces), and motivation (gamification + sponsors + future social rank) — all aimed at the same behavior: making AI-assisted coding feel measurable, competitive, and worth showing off.
If you want, I can go deeper on any one of these — scoring, the ad mechanics, or how the backend plan fits together.
hi
Hi — good to see you again.
We were looking at AI Miles (local AI usage tracking, gamification, sponsor surfaces). What would you like to do next — dig into something specific, run it, or start on a change?
what are the weaknesses of the ai miles?
Reviewing more of the codebase and docs for an honest assessment of weaknesses.
Matches: 157
Files: 78
Pattern: TODO|FIXME|brittle|impossible|caveat|honest|fabricat|cheat|legal|fragile
Path: /Users/gaganarora/Desktop/ai miles
Top files:
- ./aimiles/patch_editors.py (1)
- ./.venv/lib/python3.14/site-packages/Quartz/ImageIO/_metadata.py (1)
- ./.venv/lib/python3.14/site-packages/Quartz/CoreGraphics/_metadata.py (1)
- ./.venv/lib/python3.14/site-packages/WebKit/_metadata.py (1)
- ./browser-extension/README.md (4)
- ./browser-extension/content.js (2)
- ./COVERAGE.md (2)
- ./aimiles/wrap.py (1)
- ./aimiles/dashboard.py (1)
- ./aimiles/rank.py (2)
- ./BACKEND.md (2)
- ./.venv/lib/python3.14/site-packages/CoreFoundation/_metadata.py (1)
- ./.venv/lib/python3.14/site-packages/AppKit/_metadata.py (2)
- ./.venv/lib/python3.14/site-packages/PyObjCTools/TestSupport.py (4)
- ./.venv/lib/python3.14/site-packages/pip/_vendor/urllib3/util/url.py (1)
- ./.venv/lib/python3.14/site-packages/pip/_vendor/urllib3/util/retry.py (5)
- ./.venv/lib/python3.14/site-packages/pip/_vendor/urllib3/util/response.py (1)
- ./.venv/lib/python3.14/site-packages/pip/_vendor/urllib3/contrib/socks.py (1)
- ./.venv/lib/python3.14/site-packages/pip/_vendor/urllib3/contrib/securetransport.py (3)
- ./.venv/lib/python3.14/site-packages/pip/_vendor/urllib3/contrib/pyopenssl.py (1)Matches: 261
Files: 16
Pattern: test|pytest|unittest
Path: /Users/gaganarora/Desktop/ai miles
Top files:
- ./cursor-inject/aimiles-inject.js (1)
- ./aimiles/widget.py (1)
- ./aimiles/menubar.py (2)
- ./browser-extension/content.js (5)
- ./aimiles/wrap.py (1)
- ./.venv/lib/python3.14/site-packages/rumps/packages/ordereddict.py (1)
- ./.venv/lib/python3.14/site-packages/rumps/rumps.py (1)
- ./.venv/lib/python3.14/site-packages/Foundation/_metadata.py (3)
- ./.venv/lib/python3.14/site-packages/CoreFoundation/_metadata.py (1)
- ./.venv/lib/python3.14/site-packages/AppKit/_metadata.py (2)
- ./.venv/lib/python3.14/site-packages/objc/_transform.py (3)
- ./.venv/lib/python3.14/site-packages/objc/_pycoder.py (2)
- ./.venv/lib/python3.14/site-packages/objc/_new.py (1)
- ./.venv/lib/python3.14/site-packages/objc/_convenience_nsset.py (4)
- ./.venv/lib/python3.14/site-packages/objc/_bridgesupport.py (1)
- ./.venv/lib/python3.14/site-packages/PyObjCTools/TestSupport.py (1)Organizing AI Miles weaknesses into clear categories. The first identified issue is brittle, maintenance-heavy ad mechanics.
Ad mechanics are brittle. Binary patches break on every tool update. The PTY proxy fails with shimmer animations. Browser extensions need constant DOM upkeep. Editor patches require manual scanning per app.
Legal risks exist. Injecting ads into third-party paid products violates terms of service.
The social layer is fake. Rank and percentile estimates come from a log-normal model rather than real users. Dashboard feeds contain hardcoded placeholder content.
Backend infrastructure does not exist yet. The addiction layer remains purely speculative.
Trust and anti-cheat gaps are significant. Users can fabricate local logs. There is no attestation system. Reward redemption is local-only with no actual fulfillment.
Data coverage is incomplete. Cursor lacks token counts. Cloud agents cannot be ingested locally. JetBrains and Zed are not built. Aider requires manual configuration instead of automatic crawling. Support across tools is uneven.
Platform support is limited. The menubar relies on rumps, Cursor paths assume macOS Library folders, and installation scripts target zsh or bash.
Core value carries opt-in friction. Terminal ads require a custom alias. Users who bypass it lose both tracking and ads. Shell rc files must be modified.
Tracking and gamification clash with ad injection. Gamified users may reject ads in paid tools. Sponsors still need impressions.
There is no real test suite.
Here are the main weaknesses of AI Miles — including places where the project itself is upfront about limits.
The core monetization trick — swapping spinner text — is inherently fragile:
aimiles run -- codex) fails when tools animate spinner text per-glyph (shimmer/color per character).So a lot of the “magic” is operational debt, not a stable platform.
Injecting sponsor messages into paid third-party products (Claude Code, ChatGPT UI, VS Code core) sits in a gray area. The browser-extension README says plainly that this needs legal review before shipping. That’s a real business risk, not just a technical one.
The strongest retention hook — rank, percentile, “you passed N builders,” friends, teams — doesn’t exist yet:
estimated=True), not real users.BACKEND.md is a spec; there’s no live server, sync, or accounts.So today it feels like Strava, but the competitive/social part is simulated.
All scoring comes from local logs the user controls. BACKEND.md admits tokens can be fabricated, and mitigation (caps, anomaly detection, signed attestation) is only planned.
That’s fine for a fun personal tracker, but weak once real rewards or money are on the line. Redemption today is also local-only — redeem() deducts miles in SQLite; there’s no partner fulfillment, verification, or fraud prevention.
Tracking quality varies a lot by tool:
| Tool | Weakness |
|---|---|
| Cursor | Tokens are mostly server-side; only presence is detected, not real usage |
| Cloud agents (Devin, v0, Replit) | No local logs; browser extension is a brittle workaround |
| JetBrains / Zed | Not built yet |
| Aider | No automatic discovery; you must configure paths manually |
Power users on Cursor + cloud agents get an incomplete picture — exactly where a lot of AI coding happens.
Full value requires changing how you work:
.zshrc (codex → aimiles run -- codex)If you run tools directly, you skip both tracking and ads. Adoption depends on habit change, not passive install.
The menu-bar widget uses rumps (macOS only). Paths like Library/Application Support/Cursor and install.sh targeting .zshrc skew Mac + zsh. Linux/Windows support looks partial or absent for key surfaces.
Two value props pull in different directions:
The status-bar fallback is safer but weaker for sponsors; in-spinner ads are stronger but more hostile to users.
Code is written to be testable (menu_model(), rewrite(), etc.), but there’s no real test suite in the repo — only mentions of tests in comments and the browser extension. For something that patches binaries and parses many log formats, that’s a gap.
rewards.json lists perks (Ramp credits, Vercel credits, etc.), but there’s no evidence of live sponsor integrations, redemption flows, or inventory. It’s a catalog UI without a fulfillment backend.
Bottom line: AI Miles is strong as a clever prototype — local tracking, gamification, and creative ad surfaces. Its weaknesses cluster around sustainability: brittle integrations, simulated social proof, unproven rewards, legal exposure, and incomplete coverage of how people actually use AI today.
If you want, I can go deeper on any one area — e.g. how rank estimation works, or what would be needed to make rewards real.