Code Hangar

A local-first control centre for understanding what AI coding tools did to your projects.

Status: early alpha (0.1.x). The core navigator, discovery, safety layer and optional connector are built and tested; expect rough edges and breaking changes before 1.0. See known issues.

Claude Code, ChatGPT, Cursor, Antigravity, Hermes, OpenClaw, Pinokio and similar tools scatter projects, conversation transcripts, model files and caches across your disk. Code Hangar inventories all of it — entirely on your computer — so you can find it, understand it, navigate it, and clean it up safely.

It is not an IDE, an autonomous coding agent, a chat client or a generic disk cleaner. It is a retrospective workspace for vibe coders: reconstruct what an AI recorded, understand the affected code, and make at most one small reversible correction at a time.

---

What it does

Six things, built around local evidence:

1. Discovers your AI work. Finds the projects and the tool sessions/conversations attached to them across the AI coding apps you use — Claude Code, ChatGPT, Cursor, Antigravity, plus Hermes / OpenClaw / Pinokio — read straight from each app's own store, on Windows and (optionally) WSL, grouped by app, with the real transcripts.
2. Shows what changed. A cross-project Review Inbox finds session records newer than each saved checkpoint. Recap combines supported session edits, current local Git evidence, current-file comparison and encrypted review history. Coverage and unknowns appear before the diff; missing evidence is never invented. A private-safe Review Receipt exports counts and evidence limits without project identity, prompts, diffs, file names or paths.
3. Helps you read code. Safe source highlighting, plain-language context, Explain and What to check, bounded walkthroughs and grounded follow-up teach the reader instead of hiding the code.
4. Changes one thing safely. Project-file changes start locked and require the exact project name to unlock for that app session. Recognised values, an advanced one-file text draft and one selected AI suggestion are previewed, validated, snapshotted and reversible. Manual and value changes show the exact local line diff plus read-only Git context before applying; Code Hangar never stages, commits, pushes or changes branches. Whole-project and multi-file AI rewrites do not exist.
5. Backs up and removes safely. A journaled, reversible pipeline: nothing is destroyed without a verified backup. Files move to a holding area you can restore from; permanent removal is off by default.
6. Keeps AI optional and local-first. The Connector edition can use a local loopback model server or a provider you configure. It shows the exact request, blocks secrets before send, streams local responses and maintains an advisory per-session token meter. The Local edition contains none of this network code.

Two editions

Code Hangar ships as two editions built from the same source tree.

Edition	Who it's for	AI / network
Code Hangar (Local)	Anyone who wants the local catalog and the safe backup/remove pipeline with zero external access. The connector/AI code is physically absent from this build (CI-enforced).	none — no account, no telemetry, no outbound network
Code Hangar — AI Connector	People who also want the optional AI assist and want their AI apps to read (and, with explicit approval, act on) the curated knowledge over the Model Context Protocol (MCP).	only the local model server or provider you configure; MCP over a local channel only

The base (Local) edition never phones home. It has no outbound network capability at all — the AI and connector code is not compiled into it, which is asserted in CI. The AI Connector edition adds only the calls you opt into: an AI assist that talks to the local model server or API endpoint you choose, and an MCP channel that speaks over a local stdio / named-pipe connection — never the internet.

Even in the Connector edition, every privileged action an AI app requests is queued for your in-app approval — the AI app never executes anything itself. See docs/connect-your-ai-app.md for the full connection flow and the total-control model.

To use the review and learning layer with zero cloud, see Connect your local model.

Safety

The backup-and-remove pipeline is built to be reversible and hard to misfire:

• Backup before delete is enforced — nothing is removed without a backup that is verified to cover the file (content-hash bound).
• Reversible by default — files you clean up move to a holding area (quarantine) that you can restore from.
• Permanent removal is off by default — the one irreversible step must be explicitly enabled in Recover, and even then each removal still needs a verified backup plus a fresh confirmation. On any error the gate fails closed.
• Journaled and crash-consistent — the pipeline records what it is doing so an interrupted operation can recover rather than leave a half-finished mess.

This pipeline passed a 10-scenario adversarial battery (crash consistency, interrupted-restore recovery, holding-area collisions, backup-covers-file) with zero data loss.

Privacy & security

Code Hangar is local-first by design:

• No account and no telemetry. The Local edition has no outbound network capability. The AI Connector edition adds only the opt-in AI Assist calls to the provider or local model server you configure.
• Encrypted at rest. The local index is a SQLCipher-encrypted database, with its key bound to your Windows user account (DPAPI).
• Sensitive files stay protected — secrets, .env, key material and other Protected Zones are excluded from preview and search, and are hard-blocked before anything is sent to an AI assist.
• Support exports are redacted by construction — the diagnostics bundle contains build/safety facts and aggregate counts, never project or file identity, paths, sessions, prompts, source, diffs, logs, endpoints, credentials or model configuration.
• Destructive actions are gated — backup-before-delete is enforced; the irreversible permanent-delete is opt-in and always requires a verified backup plus a fresh confirmation.

See the public SECURITY.md policy and SECURITY_INVARIANTS.md for the detailed security model.

Install

Download the installer for the edition you want from the Releases page and run it (per-user install, no admin required):

• Code Hangar_x.y.z_x64-setup.exe — Local edition. Full local management including safe backup and delete (backup → holding → final removal, off by default); 100% local, no AI, no outbound network.
• Code Hangar AI Connector_x.y.z_x64-setup.exe — AI Connector edition. Shares the same local catalog (same app data) and adds AI Assist + the AI app integration panel.

Both editions read and write the same encrypted local catalog, so you can install either (or both) and switch later without losing your inventory.

Windows only for now. WSL projects are catalogued from Windows; there is no native Linux/macOS build yet.

Unsigned builds: until the installers are code-signed, Windows SmartScreen may warn about an "unknown publisher". Choose More info → Run anyway.

Build from source

Requires Rust stable, Node 24, and the Tauri prerequisites for Windows.

npm ci

# Connector edition installer (builds + bundles code-hangar-mcp.exe as a sidecar)
npm --workspace apps/desktop run package:connector

# Local edition installer  ->  target/release/bundle/nsis/
# Build this last if you build both editions; it cleans any staged connector sidecar.
npm --workspace apps/desktop run package:local

See docs/PACKAGING.md for the full release process and RELEASE_CHECKLIST.md for the owner's publish checklist (signing, checksums, publishing).

Development

npm ci
npm run dev            # vite dev server
npm run check          # tsc + vitest + dependency/forbidden-code guardrails
npm --workspace apps/desktop run tauri:dev   # run the desktop app

Full local gate before important pushes (this Windows machine):

powershell -ExecutionPolicy Bypass -File scripts/local-ci.ps1                 # core + mutation lanes
powershell -ExecutionPolicy Bypass -File scripts/local-ci.ps1 -AgentAutomation  # also the connector build

GitHub Actions runs the deterministic Windows gate on pushes and pull requests. The fuller local release gate above remains required before packaging. The Rust/Tauri commands need a local toolchain with cargo on PATH.

Architecture and internals: AGENTS.md, SECURITY_INVARIANTS.md, docs/, and the master spec / implementation-plan documents in the repo root.

Contributions should start with CONTRIBUTING.md. Dependency and asset provenance is recorded in SOURCES.md and THIRD_PARTY_NOTICES.md.

License

Licensed under the Apache License 2.0.