jawata-studio

JAWATA Studio

The desktop control plane for JAWATA — surgical, risk-free Java refactoring for autonomous agents. jawata-studio downloads the JAWATA engine, manages named workspaces of Java projects, runs them efficiently, and wires JAWATA into your AI agent's MCP config in one click.

A Tauri desktop app for Linux, macOS, and Windows. You point it at your Java projects; it does the rest.

# Linux — one line installs the app and a desktop entry
curl -fsSL https://jawata.org/install.sh | bash

---

Install

Platform	How
Linux (x86_64 / aarch64)	curl -fsSL https://jawata.org/install.sh | bash — downloads the matching .AppImage and registers a desktop entry. Or grab the .AppImage / .deb from releases.
macOS (Apple Silicon)	Download the .dmg from releases and drag to Applications. Unsigned: right-click → Open once to clear Gatekeeper.
Windows (x64 / ARM64)	Download the .msi or _x64-setup.exe from releases. Unsigned: on the SmartScreen prompt choose More info → Run anyway.

Prerequisite: Java 21+ on the PATH — the JAWATA engine it runs is a JVM process.

The installer pulls the right artifact for your CPU automatically. To update, re-run the one-liner (Linux) or install the newer package; jawata-studio also self-updates the JAWATA engine from its releases.

---

What it does

• Workspace-first. Group Java projects into named workspaces. Each workspace loads into one JAWATA process and is exposed as a single MCP service your agent can call. Add a project by browsing to it, or point autoscan at a parent folder (e.g. ~/Projects) to import a whole tree.
• One JVM per workspace, not per client. jawata-studio runs a single resident JAWATA engine per workspace and writes a URL endpoint into each client config. Three agents and a workspace are one JVM, not one per agent — memory stays flat no matter how many clients connect.
• One-click MCP deploy. Writes the correct server entry into Cursor, Claude Desktop, Antigravity, and IntelliJ-style configs — each in that client's own schema — and re-syncs them when a workspace changes.
• Makes the agent actually use JAWATA. Deploy writes a trigger→tool guide (a "prefer JAWATA over grep" rule with an intent→tool table) into each client and — on Claude Code — a PreToolUse hook that enforces try-first: a grep over .java, or a hand-edit of an existing .java file, is blocked with a redirect to the right JAWATA tool — unless you already looked it up via JAWATA this session, or you declare jawata-fallback: <why> (logged, versioned). Advertising the tools wasn't enough; the hook makes not using them the inconvenient path. Health-gated (engine down → ask, don't silently degrade); non-Java work is left untouched.
• Learns from what happens next. A companion PostToolUse observer (never blocks) records declared fallbacks, ungrounded .java reads, and compile/test outcomes into a versioned log — the signal for sharpening the guidance over time, and JAWATA's own feature backlog.
• Auto-managed engine. Polls for new JAWATA releases and downloads the matching runtime; the jar path stays stable across updates.
• Lives in the tray. A system-tray menu drives per-workspace start/stop without opening the window. Optional autostart-on-boot restores the workspaces you had running.
• One memory, every client. The JAWATA knowledge store is your agent's durable, cross-client memory — markdown files in your git plus a symbol-anchored store, written and recalled from Cursor, Claude Code, or any MCP client. In Claude Code it's pushed into sessions automatically (primer + recall hooks); in Cursor, ask: "what do we know about freeSlot?" (recall), "record this as a lesson anchored to pipeline.SlotManager" (record), "prime yourself from the memory store" (session primer). Or skip the agent entirely: any markdown is memory — and your docs/ folder is probably already a corpus. Add it under Memory sources and Load turns sprint docs, postmortems and ADRs into agent knowledge: one fact per section, search cues from the headings, auto-anchored to the code the text names — the next agent touching freeSlot gets handed the postmortem written about it. If an agent says "I'll remember that" without calling JAWATA, it's remembering into an opaque chat memory no other tool can see — durable memory is the one in the studio's Memory view: listable, promotable, exportable, yours.

---

How it works

jawata-studio is a pure orchestrator — a Tauri (Rust + web UI) app. It never analyses Java itself; that is entirely the JAWATA engine's job. The manager:

1. Resolves and downloads the platform-matched jawata-mcp engine.
2. For each workspace, writes <data-dir>/workspace.json (the project list the engine watches) and launches one resident engine on an allocated (port, token).
3. Deploys the resulting MCP endpoint into your chosen clients' configs and keeps them in sync.

State lives under your platform's standard config / state / cache directories for jawata-studio.

System tray on Linux

The tray icon uses AppIndicator. On GNOME, install the AppIndicator extension so the icon and its menu appear; KDE and most other desktops show it out of the box. Left-click opens the menu (status glyphs: ● running · ◐ starting · ○ stopped · ✗ failed).

---

Building from source

git clone https://github.com/haraldwegner/jawata-studio.git
cd jawata-studio
npm install
npm run tauri build          # produces installers under src-tauri/target/release/bundle

Requires the Tauri prerequisites (Rust + your platform's webview/build deps) and Node.js. npm run tauri dev runs the app against a dev build.

---

License

AGPL-3.0 — see also NOTICE. jawata-studio is the desktop control plane for JAWATA. Contributions are accepted under the Contributor License Agreement; see CONTRIBUTING.md.