The Problem

AI coding assistants are powerful but undisciplined. Ask them to build a login form and you'll get a login form plus an auth controller, a database schema, and a session handler you never asked for.

This creates tangled codebases where files are owned by no one, logic lives in the wrong layer, and integration becomes a debugging session.

The Solution

CapDrop gives your AI a job title and a contract. Each Capsule is a strict, role-specific system prompt that locks the AI into one lane — and when it hits the edge of its role, it stops and tells you exactly what needs to happen next instead of quietly doing someone else's job. Drop UI/UX capsule → AI builds frontend only Drop Backend capsule → AI builds server + database only Drop QA capsule → AI writes tests only, never touches production code

How It Works

flowchart TD
    A([Developer opens workspace]) --> B[CapDrop activates on startup]
    B --> C[Local WebSocket server starts on :5050]
    C --> D{How do you want to shift capsule?}

    D -->|Command Palette| E[CapDrop: Open Capsule Library]
    D -->|Tauri Desktop Widget| F[Click capsule card in floating overlay]

    E --> G[Select capsule from library grid]
    F --> G

    G --> H[Capsule system prompt compiled]
    H --> I{PRD.md / TRD.md in workspace root?}

    I -->|Yes| J[Specs appended as anchored constraints]
    I -->|No| K[Capsule boundary rules applied standalone]

    J --> L[Full system prompt injected into AI session]
    K --> L

    L --> M[intent_log.json written to .gemini/]
    M --> N[AI executes within capsule boundary]
    N --> O{AI hits boundary of its role?}

    O -->|No| P[Task completed in lane ✅]
    O -->|Yes| Q[CAPSULE BOUNDARY REACHED message output]
    Q --> R[Developer switches to correct capsule]
    R --> G

Capsules

Every capsule enforces a hard boundary. When the AI reaches the edge of its role it outputs: CAPSULE BOUNDARY REACHED: This task requires backend work. Please switch to the Backend capsule to complete: [one sentence description]. The frontend scaffold has been built to the extent possible.

Architecture

graph TB
    subgraph VSCode ["VS Code Extension Host"]
        EXT["extension.ts\nActivation + Auth"]
        SERVER["server.ts\nWebSocket :5050"]
        RUNNER["runner.ts\nSession Runner"]
        PROMPTS["prompts.ts\nAll Capsule Definitions"]
        LOGGER["logger.ts\nintent_log.json writer"]
        VALIDATOR["validator.ts\nIntent Gate"]
        TERMINAL["terminal.ts\nSecure Build Runner"]
        INTEGRATION["integration.ts\nBuild Repair Loop"]
        LIBRARY["library.ts\nWebview Panel"]
    end

    subgraph Tauri ["Tauri Desktop Widget (Optional)"]
        APP["App.tsx\nCapsule Cards UI"]
        RUST["main.rs\nWS Client + Session ID"]
    end

    subgraph Workspace ["Local Workspace"]
        PRD["PRD.md (optional)"]
        TRD["TRD.md (optional)"]
        LOG[".gemini/intent_log.json"]
    end

    PROMPTS --> RUNNER
    SERVER -->|persona_shift event| RUNNER
    RUNNER --> LOGGER
    RUNNER --> VALIDATOR
    RUNNER --> TERMINAL
    TERMINAL --> INTEGRATION
    LIBRARY -->|team_sync event| SERVER
    APP --> RUST
    RUST -->|WebSocket message| SERVER
    PRD --> RUNNER
    TRD --> RUNNER
    LOGGER --> LOG

Capsule Session State Machine

stateDiagram-v2
    [*] --> Idle : Extension activates on startup

    Idle --> CapsuleSelected : Developer selects capsule
    CapsuleSelected --> LoadingSpecs : Check workspace for PRD.md / TRD.md

    LoadingSpecs --> CompilingPrompt : Specs found — append to capsule prompt
    LoadingSpecs --> CompilingPrompt : No specs — use capsule boundary rules only

    CompilingPrompt --> SessionActive : Prompt compiled and injected
    SessionActive --> IntentLogged : intent_log.json written

    IntentLogged --> ValidationGate : Supabase validator called (if configured)
    ValidationGate --> Executing : Approved
    ValidationGate --> Blocked : Rejected
    ValidationGate --> Executing : Developer override

    Blocked --> Idle : Session aborted

    Executing --> BoundaryReached : AI hits capsule edge
    Executing --> Completed : Task finished in lane

    BoundaryReached --> Idle : Developer switches capsule
    Completed --> Idle : Ready for next capsule

Project Structure

capdrop/ ├── extension/ # VS Code Extension (TypeScript) │ ├── src/ │ │ ├── extension.ts # Activation, auth, command registration │ │ ├── server.ts # Local WebSocket server :5050 │ │ ├── runner.ts # Capsule session compiler and runner │ │ ├── prompts.ts # ALL capsule definitions live here ← │ │ ├── logger.ts # intent_log.json writer │ │ ├── validator.ts # Intent validation gate │ │ ├── terminal.ts # Secure build command executor │ │ ├── integration.ts # Build repair loop │ │ ├── library.ts # Capsule Library webview panel │ │ └── http.ts # Native Node.js HTTP/HTTPS client │ ├── icon.png # Extension icon (128x128) │ ├── package.json # Extension manifest │ ├── README.md # Marketplace README │ └── LICENSE # AGPL v3 │ ├── tauri-capsule/ # Desktop Widget (Rust + React, optional) │ ├── src/ │ │ ├── App.tsx # Capsule card UI │ │ └── main.tsx # React entry │ ├── src-tauri/ │ │ ├── src/main.rs # Rust WS client + session ID generation │ │ └── tauri.conf.json # Borderless, always-on-top window config │ └── package.json │ └── supabase/ # Optional cloud validator (self-hosted) ├── functions/ │ └── intent-validator/ │ └── index.ts # Edge function: validates intent_log vs specs └── migrations/ └── 20260608000000_init.sql

Installation

From the Marketplace

Search CapDrop in the VS Code Extensions panel or install via:

code --install-extension antigravity.capdrop

From Source

git clone https://github.com/your-repo/capdrop
cd capdrop/extension
npm install
npm run deploy

Commands

Optional: Tauri Desktop Widget

The floating desktop widget shows your active capsule and lets you switch roles without opening VS Code's command palette.

cd tauri-capsule
npm run tauri dev

Requires Rust installed. The widget communicates with the extension over ws://localhost:5050.

Optional: Spec Anchoring

Drop a PRD.md and/or TRD.md in your workspace root. CapDrop appends them automatically to the active capsule's system prompt, giving the AI both role boundaries and project-specific constraints.

No specs required — capsule boundaries work standalone.

Privacy

• Source code never leaves your machine
• intent_log.json is local only
• No telemetry, no analytics, no accounts required
• No API keys — CapDrop injects prompts into your existing AI assistant

Contributing

Want to contribute a capsule, fix a bug, or suggest a feature?

See CONTRIBUTING.md

All capsule definitions live in extension/src/prompts.ts as entries in ALL_PERSONAS. One PR = one capsule.

License

This project is licensed under the GNU Affero General Public License v3.0 (AGPL-3.0).

You are free to use, study, and contribute to this project. Any commercial use, distribution, or derivative work must also be released under AGPL-3.0. For commercial licensing inquiries, open an issue or contact the maintainers directly.

See the LICENSE file for full terms.