Repository navigation
CodeTwin ML v1.0.0 — Initial Release
A local-first desktop app that builds a Software Digital Twin of your codebase, finds quality and security problems, and helps you fix them safely — with proof.
Released: 2026-11-18 · Platform: Windows x64
Core rule: CodeTwin records only what it can prove. Anything missing, skipped, or uncertain is shown as such — it is never quietly turned into a "pass".
What is CodeTwin ML?
CodeTwin ML runs entirely on your machine. Point it at a local repository and it builds a Software Digital Twin — a persistent, searchable model of the project's files, symbols, dependencies, findings, and evidence. On top of that twin it layers code-quality, security, database, runtime, and QA analysis, optional on-device ML, authorized web-security testing, and hash-verified repair workflows.
Nothing is uploaded. No model is bundled. Your code never leaves the device.
What it does
🧠 Understand any codebase
- Detects languages, frameworks, package managers, test setups, databases, and CI.
- Indexes files, functions, classes, imports, and references with Tree-sitter for TypeScript/JS, Python, Rust, Go, C, C++, and PHP.
- Every supported file is SHA-256 hashed; re-indexing only re-parses what changed.
- Impact analysis: see a file or symbol's dependencies, dependents, and reverse "what would this break" blast radius.
- Optional LSP-backed semantic evidence. An unresolved import is shown as unresolved — never as a made-up link.
🔎 Find real problems (static, no code executed)
- Code quality — oversized definitions, deep nesting, high fan-out, dependency cycles.
- Static security — credential-like literals (redacted, values never stored), dynamic code execution, weak hashes, unsafe C/C++ string APIs, risky request flows.
- Dependency audit — known-vulnerable dependencies via OSV.
- Database — destructive DDL, unscoped
UPDATE/DELETE, disabled SQLite foreign keys, literal Prisma datasource URLs. It never connects to your database. - Runtime config — Dockerfile / Compose reproducibility and health/restart risks. It never starts containers.
- Cross-language security graph — FFI, REST, gRPC, and IPC boundaries plus source→sink taint tracing, with a strict evidence ladder (structural signals are never promoted to "confirmed").
🧪 QA & test evidence
- Passive discovery of test frameworks, test files, and config — discovery never implies the tests ran.
- On Windows, explicitly selected tests can run inside a zero-capability sandbox with Job Object limits. A verdict is recorded only when the runner finished and its output was understood — timeouts or crashes give no verdict, not a pass.
🌐 Authorized web-security testing
- Guided, scope-limited active testing against sites you are authorized to test, with an explicit verification step.
- Findings link back to the source code that produced them.
🛡️ Security operations & triage
- Static malware triage — byte-only inspection (never executes a sample): format, entropy, strings, IOCs, capabilities, and signatures → a graded verdict, bridged into the incident/quarantine lifecycle.
- Endpoint-security and SOC event handling.
🤖 On-device AI (optional)
- Runs integrity-checked local ONNX / GGUF models through a bundled llama.cpp runtime — assistive code review, a project-grounded workspace chat, and a Setup Agent that can connect a local model and download security datasets (CWE/CAPEC/ATT&CK/Exploit-DB/SecLists).
- No model ships in the installer; nothing is uploaded.
🔧 Fix safely, with proof
- Proposes a small patch, shows the exact diff, applies it only after you approve, backs up originals, and re-tests to confirm the fix.
Why it's different (key advantages)
| Advantage | What it means for you |
|---|---|
| Local-first & private | All indexing, analysis, and ML run on your machine. Code and secrets never leave the device. |
| Evidence-based, no fake passes | Uncertain or skipped checks are reported honestly — the twin stores only what it can prove. |
| Safe by construction | Analysis reads source; it never runs your app, connects to your DB, or starts containers. Fixes require your approval and keep backups. |
| One twin, many analyzers | Quality, security, database, runtime, QA, and cross-language boundary analysis all share one indexed model with linked evidence. |
| Cross-language depth | 7 parsed languages plus FFI/REST/gRPC/IPC boundary and taint analysis across them. |
| Tunable performance | Eco / Balanced / Maximum CPU modes (CPU-affinity capped), switchable live without restart. |
| Workspace built for teams | Organization profile, per-project AI chat, review stat bar, and user profile image in sidebar/header. |
What's in this release
| Artifact | Platform | Notes |
|---|---|---|
CodeTwin ML_1.0.0_x64-setup.exe |
Windows x64 | NSIS installer (per-machine) |
CodeTwin ML_1.0.0_x64_en-US.msi |
Windows x64 | WiX / MSI installer |
Launch-CodeTwin-ML.bat |
Windows | Launches the app, or the installer if not yet installed |
run-codetwin-ml.sh |
Linux | Launcher for a locally produced Linux build |
CodeTwin-ML.command |
macOS | Launcher for a locally produced macOS build |
metadata/SHA256SUMS.txt |
— | SHA-256 of every artifact above |
Install (Windows x64)
- EXE: double-click
CodeTwin ML_1.0.0_x64-setup.exe. - MSI: double-click it, or
msiexec /i "CodeTwin ML_1.0.0_x64_en-US.msi".
Verify your download
Get-FileHash ".\windows\x64\CodeTwin ML_1.0.0_x64-setup.exe" -Algorithm SHA256Compare the result against metadata/SHA256SUMS.txt.
Platform support & known limitations
- Windows x64 is the only shipped binary target for v1.0.0, and it requires 64-bit Windows.
- Windows x86 (32-bit) is not provided. The bundled browser sidecar embeds the host's x64 Node + Chromium and does not cross-compile, so a genuine 32-bit build cannot be produced.
- Linux (
.sh) and macOS (.command) ship as launcher scripts, not binaries. Tauri apps compile on their target OS and cannot be cross-built from Windows. To get a native build, runcd apps/desktop && npm run tauri:buildon that OS, then place the produced AppImage/.deb/.appnext to the matching launcher. - On-device ML and web-security testing are optional and off until you enable them. The sandboxed test runner is Windows-only.
First run
- Add a project — Dashboard → Add Project Folder — pick a local repository. It is indexed without running any of its code.
- Explore the Digital Twin under Code Analysis (files, symbols, dependencies, impact).
- Analyze — run quality, security, database, and runtime checks; review each finding with its evidence.
- Test / security-test / fix — optional, opt-in, and always on targets you own or are authorized to test.
CodeTwin ML — understand your code, find what's wrong, fix it with proof. All on your machine.