Stable release
v1.0.0 — v1.0.0 - Initial Relesase
November 18, 2026 · 6 published files
Read release description
# 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
```powershell
Get-FileHash ".\windows\x64\CodeTwin ML_1.0.0_x64-setup.exe" -Algorithm SHA256
```
Compare 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, run `cd apps/desktop && npm run tauri:build` on that OS, then place the produced AppImage/`.deb`/`.app` next 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
1. **Add a project** — *Dashboard → Add Project Folder* — pick a local repository. It is indexed without running any of its code.
2. **Explore** the Digital Twin under **Code Analysis** (files, symbols, dependencies, impact).
3. **Analyze** — run quality, security, database, and runtime checks; review each finding with its evidence.
4. **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.*