AUDIT
EVERYTHING
YOU'VE BUILT.
Point it at a folder of projects. An on-device model grades every one — missing pieces, stale deps, bloat, UI rot — then turns your profile into project ideas worth your weekend. Nothing ever leaves your machine.
TWO INSTRUMENTS.
ONE SHELL.
Run either alone or together. Both driven entirely by a model living on your disk.
Project Auditor
Aim it at a root folder. Every subfolder becomes a project, gets typed from its marker files, and graded across five axes — each finding ranked critical suggested optional.
Idea Generator
Fill an About Me — skills, tools, goals, free time, domains — and get a categorized list of personal builds. Every idea carries effort and impact, favoritable and re-rollable.
FIVE AXES
PER PROJECT.
Each project is dissected on five fronts. Every finding sorted, colored, and ranked by severity.
Missing pieces
What a project of this type should ship but doesn't — README, tests, license, CI, error handling.
Feature ideas
Additions that fit what the project actually is — each with a rationale and an effort hint.
Outdated tech the only networked axis · opt-in
Stale languages, frameworks, and libraries, shown as installed vs recommended.
UI assessment attach a screenshot
Source alone can't judge a UI. Drop in a screenshot and the model critiques modernity, spacing, contrast.
Bloat
Files, dependencies, and features that are dead weight and safe to remove.
POINT.
SCAN. READ.
Two flows, both fully on-device. The network is touched once — to check dependency versions — and only if you allow it.
Pick the root that holds all your projects.
The on-device model grades each project across five axes.
Per-project deep dives plus a ranked portfolio action list.
Skills, tools, goals, time, and the domains you care about.
The model drafts ideas that fit exactly who you are.
Filter by domain, sort by effort, keep the best.
YOUR CODE
NEVER LEAVES
THIS MACHINE.
No account. No sign-in. No servers to trust. Codescry runs a language model locally and stores every result on disk — yours alone.
Install and run. There is nothing to sign up for, ever.
The app sends no usage data, ever, and reaches no server of ours. This website counts visits — cookieless and aggregate.
Fully functional in airplane mode. Disconnect and nothing changes.
▸ Two exceptions, both opt-out with one switch: the Outdated-tech axis can compare your dependency versions against public registries, and the app asks GitHub once per launch whether a newer release exists. Neither sends code or usage data — and both stop the instant you flip “Go fully offline.” This page, like the rest of codescry.app, keeps a cookieless visitor count that honours Do Not Track and Global Privacy Control; the app does not. Full detail ▸
A SHARP
INSTRUMENT.
Dense, precise, dark by default. Built for a developer who lives in the terminal — not a toy.
Add .github/workflows/ci.yml to run tests on every push.
Two major versions behind. Migration guide available.
214 MB reclaimable · not referenced in source.
A CLI that fills missing FLAC metadata on-device.
Ping your Pi services, show status on a spare monitor.
Summarize commits into release notes, fully local.
Pomodoro that silences Slack and DNDs your machine.
GET
CODESCRY.
One binary. No installer account, no license key. Free, with an open local model. Pick your platform — your detected OS is marked.
These bundles carry no OS code signature yet. Windows SmartScreen flags the .msi as an unrecognised app, and macOS Gatekeeper refuses the .dmg until you allow it once under System Settings → Privacy & Security. Nothing is wrong with the build — check it against SHA256SUMS on the release page, which is the integrity proof we do publish today.
the model runs in memory
model + local results store
falls back to CPU, just slower
STRAIGHT
ANSWERS.
The seven things people ask before installing. No marketing hedge — this is how it actually works.
01 / MODEL Which model does it run, and how big is it?
An open model served by a local Ollama instance on 127.0.0.1:11434 — nothing hosted, nothing keyed. The default is gemma3:4b (~3.3 GB), which also reads screenshots, so the UI axis works. You can switch to qwen3:4b (~2.6 GB, stronger code reasoning, no vision), llama3.2:3b (~2 GB, the fastest and lightest), or gemma3:12b (~8.1 GB, slower but sharper). Pulled once, then it lives on your disk.
02 / OFFLINE Does it work offline?
Yes — completely, once the model is pulled. Scanning, scoring, reports, heatmap, and idea generation all run in airplane mode. Codescry makes exactly two kinds of outbound request: opt-in dependency version lookups against public registries, and a check-only update ping. Flip “Go fully offline” and both are skipped entirely — not queued, not retried.
03 / PRIVACY Is my code uploaded anywhere?
No. Your source, screenshots, and results never leave the machine. There is no account, no server, and no telemetry in the app — there is nowhere for it to go. The registry check, when you enable it, sends package names and version numbers only, never file contents. This website (not the app) keeps a cookieless visitor count; it never sees anything you scan.
04 / PROJECTS Which project types are supported?
Each subfolder is typed from its marker files: node (package.json), python (pyproject.toml, requirements.txt, setup.py, Pipfile), rust (Cargo.toml), .net (.csproj, .sln, .fsproj), godot (project.godot), go (go.mod), and java (pom.xml, build.gradle). Anything unrecognized is still scanned as generic — same five axes, a lighter list of expected files. Registry version checks cover npm, PyPI, and NuGet.
05 / HARDWARE What are the minimum hardware requirements?
8 GB RAM to run comfortably (16 GB recommended), ~4 GB free disk for the model plus the local store, on Windows 10/11 64-bit, macOS (Apple Silicon or Intel), or a modern Linux. A GPU or NPU is optional — Codescry falls back to CPU and simply takes longer per project. The small models need ~4–6 GB of headroom; gemma3:12b wants ~12 GB.
06 / UPDATES How do I update?
Codescry checks the public releases page and tells you when a newer version exists — it never downloads or installs anything on its own. You open the release page, grab the build for your platform, and install it like any other app. The check is skipped in offline mode, and your local database, settings, and history survive the upgrade untouched.
07 / STORAGE Where are my results stored?
In one SQLite file — codescry.db — inside the app-data directory: %APPDATA%\org.helveticlabs.codescry\ on Windows, ~/Library/Application Support/org.helveticlabs.codescry/ on macOS, ~/.local/share/org.helveticlabs.codescry/ on Linux. Reports and the priority list export to .md or .json wherever you point them. Delete that folder and every trace of Codescry's data is gone.
▸ Still unsure? The docs go deeper, and support@codescry.app is a real inbox — no ticket queue.