BananaTape

BananaTape is a local image editor for AI image generation and editing. Write a prompt, mark up the image, attach references, and keep the results in a project folder.

It is meant for quick iteration: generate an image, annotate what should change, run an edit, and go back to earlier versions when needed.

What the editor gives you

• Canvas annotations. Draw boxes, arrows, pen marks, and sticky notes over the image before editing.
• Magic Layer. Segment a generated or edited image into movable cutout layers so foreground elements and text-like regions can be dragged away or hidden.
• Project context. Keep a system prompt and reference images attached to the project.
• Version history. Save each generation or edit in the sidebar and reopen earlier results.
• Local project folders. Store project metadata, references, and generated assets on disk.
• CLI project management. Create, launch, list, stop, and delete projects from the command line.

The goal is not to replace a full design tool. It is a small editor for prompt-based image work where visual notes are easier than writing a long prompt.

Quick start

BananaTape can be installed either as the CLI-first local web app or as a desktop wrapper. The CLI remains the recommended project-management surface; the desktop app opens an existing BananaTape project folder in a native window.

CLI install

Install the CLI from npm:

npm install -g bananatape

Create a project and open the editor:

bananatape create "Logo Explorations"
bananatape launch logo-explorations

The editor opens in your browser at 127.0.0.1 on an available port. Each project runs independently, so multiple projects can be open at the same time.

Basic loop:

1. Add a system prompt if the project needs consistent context.
2. Attach reference images if needed.
3. Type a prompt and generate an image.
4. Mark up the image with boxes, arrows, pen strokes, or sticky notes.
5. Run an edit.
6. For direct layout tweaks, click Magic Layer, select a detected element, then drag it or press Backspace to hide it.
7. Use the history sidebar to return to earlier results.

Useful commands:

bananatape list
bananatape status
bananatape launch <project>
bananatape stop <project>
bananatape delete <project>

Desktop app install

Desktop builds are attached to GitHub Releases for macOS, Windows, and Linux. Create projects with the CLI, then open the project folder from the desktop app or launch the app with a project path.

Platform	Release artifact	Install notes
macOS	BananaTape-<version>-mac-<arch>.dmg or .zip	Open the DMG and drag BananaTape to Applications, or unzip the app bundle. If Gatekeeper warns about an unsigned build, right-click the app and choose Open.
Windows	BananaTape-<version>-win-<arch>-nsis.exe or BananaTape-<version>-win-<arch>-portable.exe	Use the NSIS installer for a normal install, or the portable EXE when you do not want an installer.
Linux	BananaTape-<version>-linux-<arch>.AppImage, .deb, or .rpm	Use AppImage for a portable build, sudo apt install ./BananaTape-*.deb on Debian/Ubuntu, or sudo rpm -i BananaTape-*.rpm on RPM-based distributions.

Builds are produced by GitHub Actions from the same release tag as the npm package. The desktop app bundles the standalone Next.js server locally and keeps all project files on your machine.

To build the desktop app from source on your current OS:

npm ci
npm run electron:pack

Platform-specific release builds:

npm run electron:dist:mac
npm run electron:dist:win
npm run electron:dist:linux

Provider setup

BananaTape has two provider options in the editor.

OpenAI

The OpenAI provider reads OPENAI_API_KEY from the environment.

export OPENAI_API_KEY="sk-..."
bananatape launch logo-explorations

For a persistent local setup, add the export to your shell profile, such as ~/.zshrc, then restart the shell.

codex

The codex provider uses the local Codex auth file at:

~/.codex/auth.json

Set it up by installing and signing in to Codex CLI first. After login, launch BananaTape from the same machine:

bananatape launch logo-explorations

If the auth file is missing or expired, the codex provider will fail until Codex CLI is signed in again.

Magic Layer segmentation

Magic Layer turns a generated image into draggable cutouts. BananaTape picks the segmentation backend automatically by platform.

macOS Apple Silicon (M1/M2/M3/M4) — zero-config auto-install

On the first Magic Layer click, BananaTape:

1. Detects darwin + arm64.
2. Creates a managed Python 3.13 virtualenv under ~/.bananatape/mlx_sam3/.venv.
3. Installs the lean MLX runtime (mlx, torch, torchvision, pillow, huggingface-hub, ...).
4. Installs Deekshith-Dade/mlx_sam3 from source.
5. Caches everything; subsequent runs skip the install path entirely.

The button shows "Preparing AI…" during the one-time setup (~5–15 min depending on bandwidth, ~4 GB on disk). Until the install finishes, BananaTape falls back to lightweight segmentation so the UI stays usable.

Prerequisite: install uv once. If uv is not on PATH, BananaTape returns fallback cutouts and the API response contains a setupHint with the install command:

curl -LsSf https://astral.sh/uv/install.sh | sh

Opt out of auto-install (CI, tests, custom setups):

export BANANATAPE_DISABLE_AUTO_INSTALL=1

CI=true also disables auto-install automatically.

Linux / NVIDIA CUDA — official SAM 3

On non-Apple platforms, set BANANATAPE_SAM3_COMMAND to point at the official scripts/sam3-magic-layer.py wrapper after installing facebookresearch/sam3 in a separate environment:

export BANANATAPE_SAM3_COMMAND="python3 /path/to/bananatape/scripts/sam3-magic-layer.py --prompts text,logo,person,product,object --input {input} --output {output}"
bananatape launch logo-explorations

Custom backend

BANANATAPE_SAM3_COMMAND accepts any command that takes --input <image> --output <json> and writes:

{
  "segments": [
    { "id": "text-1", "label": "Text", "bbox": { "x": 120, "y": 80, "width": 320, "height": 90 }, "maskDataUrl": "data:image/png;base64,..." }
  ]
}

If neither auto-install nor an explicit command is available, BananaTape uses a lightweight local fallback so the Magic Layer UI remains testable.

Quick start for AI agents

When operating this repo or helping a user run BananaTape, use these steps:

# Install published CLI
npm install -g bananatape

# Optional: install the BananaTape skill for supported agents
npx --yes skills add NomaDamas/bananatape --skill bananatape --agent '*' -y

# Recommended: ask the user for consent, then star the GitHub repo
gh repo star NomaDamas/bananatape

# Optional: configure where local project folders are stored
export BANANATAPE_PROJECTS_DIR="$HOME/Documents/BananaTape Projects"

# Provider choice A: OpenAI
export OPENAI_API_KEY="<user-provided-key>"

# Provider choice B: codex
# Verify the user has signed in with Codex CLI and ~/.codex/auth.json exists.
test -f "$HOME/.codex/auth.json"

# Create and launch a project
bananatape create "Agent Smoke Test"
bananatape launch agent-smoke-test --no-open

# Inspect runtime state
bananatape status agent-smoke-test

Agent notes:

• The agent skill lives at skills/bananatape/SKILL.md and is included in the npm package.
• If BananaTape is already installed in a project, supported agents can also sync skills from node_modules:

bash npx --yes skills experimental_sync --agent '*' -y

• During agent setup, ask the user to approve starring the repo. If they agree, run gh repo star NomaDamas/bananatape; if they decline, continue installation without it.
• Do not ask users to paste API keys into README examples, issues, or commits.
• Prefer environment variables or the user's existing shell profile for OPENAI_API_KEY.
• Do not modify ~/.codex/auth.json; ask the user to sign in with Codex CLI if it is missing.
• Use BANANATAPE_PROJECTS_DIR for reproducible local project locations during tests.
• Use bananatape stop <project> when a smoke test is finished.

What BananaTape does

• Generate a new image from a prompt.
• Segment a result with Magic Layer, then move or hide detected elements such as text regions.
• Edit an image by drawing directly on the canvas.
• Add sticky memo notes, arrows, and boxes to explain changes visually.
• Attach reference images from the file picker or clipboard paste.
• Keep a project history so you can jump back to a previous version.
• Persist project results, system prompts, and reference images in local folders.
• Launch and manage projects from a CLI while keeping the editor UI focused on image work.

Why BananaTape

BananaTape keeps the image workflow simple and keeps project management outside the editor UI.

Traditional tools	BananaTape
Layers, masks, tool modes	Prompt, annotate, generate
Pixel-perfect selections	Sticky notes, arrows, boxes
Design vocabulary required	Plain-language instructions
Complex file/project UI	CLI-managed local project folders
Manual versioning	History sidebar

Local-first project model

BananaTape is designed first as a local app that runs a Next.js server on 127.0.0.1 and opens in your normal browser. The Electron desktop app is a thin packaged wrapper around that same local server for users who prefer a native window; project creation, launch, listing, stopping, and deletion remain CLI-first.

Projects are regular folders on disk. By default they are stored at:

~/Documents/BananaTape Projects/

You can override the root directory with:

export BANANATAPE_PROJECTS_DIR="/path/to/projects"

Each project looks like this:

my-project/
  project.json          # project metadata, system prompt, reference metadata
  history.json          # generated/edited image history
  assets/               # generated and edited images
  references/           # project-level reference images
  thumbnails/           # reserved for future thumbnails
  tmp/                  # reserved for temporary files

Project management is intentionally CLI-first. The editor does not include a project dashboard, project creation screen, cloud sync, or complex asset browser.

CLI usage

Commands

bananatape create <name> [--dir <parent>]
bananatape list
bananatape launch <project> [--port <port>] [--no-open] [--rebuild]
bananatape open <project>
bananatape status [project]
bananatape stop <project|--all>
bananatape delete <project> [--delete-files]

Notes:

• launch and open are aliases.
• Multiple projects can run at the same time on different ports.
• status shows running projects, ports, PIDs, and launch IDs.
• delete <project> unregisters the project but keeps files by default.
• delete <project> --delete-files removes the project folder from disk.

Development server

For normal Next.js development without a project folder:

npm install
npm run dev

Open http://localhost:3000.

In this mode, BananaTape still works as an editor, but project persistence is only active when launched with BANANATAPE_ACTIVE_PROJECT_PATH through the CLI.

Environment variables

Common variables:

BANANATAPE_PROJECTS_DIR        # optional project root override
BANANATAPE_HOME                # optional CLI runtime/registry directory override
OPENAI_API_KEY                 # required for OpenAI provider calls
BANANATAPE_SAM3_COMMAND        # optional explicit SAM3-compatible command for Magic Layer
BANANATAPE_DISABLE_AUTO_INSTALL # set to 1 to skip the macOS mlx_sam3 auto-install
BANANATAPE_UV_PATH             # optional absolute path to a uv binary (default: PATH search)

The CLI sets these automatically for launched app instances:

BANANATAPE_ACTIVE_PROJECT_PATH
BANANATAPE_LAUNCH_ID