Open Source · Apache-2.0 · Rust v0.9.8

Multimodal preflight
for AI APIs. Automatic.

SHIFT sits between your app and the model API. It inspects, evaluates, and transforms images so every request is valid, optimized, and tuned to your cost/quality preference.

60-90% size reduction

<10ms overhead

8+ image formats

Install SHIFT Star on GitHub

shift-ai

The problem with raw image payloads

AI model APIs have strict constraints on image inputs. Violating them wastes tokens, causes hard failures, or crashes sessions.

📷

Oversized images

A 4000x3000 screenshot dropped into a prompt consumes massive tokens. Anthropic charges per-pixel; OpenAI uses tiles. Either way, you're paying for resolution the model can't use.

~1,568 tokens wasted

⚠️

Hard 400 errors

Send an SVG, BMP, or TIFF to a model API and you get a 400 error. The API rejects the request outright. Your agent retries, wastes more tokens, and often gives up.

Session crash

💰

Token cost bloat

Multimodal tokens are expensive. Without optimization, a 30-min coding session with screenshots can burn through 50-100K tokens on images alone. Most of that is pure waste.

$2-5 per session

Before & After

See what SHIFT does to a real API request payload.

Without SHIFT ~42 KB payload

{
  "model": "gpt-4o",
  "messages": [{
    "role": "user",
    "content": [{
      "type": "image_url",
      "image_url": {
        "url": "data:image/png;base64,iVBORw0KGgo..."
      }
    }]      // 4000x3000 PNG, 42KB
  }]        // 1,568 Anthropic tokens
}           // Exceeds provider limits

With SHIFT (economy) ~17 KB payload -60%

{
  "model": "gpt-4o",
  "messages": [{
    "role": "user",
    "content": [{
      "type": "image_url",
      "image_url": {
        "url": "data:image/png;base64,iVBORw0KGgo..."
      }
    }]      // 1024x768 PNG, 17KB
  }]        // 1,082 Anthropic tokens (-31%)
}           // All constraints met

Four-stage pipeline

Every request passes through four stages. Each image is inspected, evaluated against provider rules, transformed, and slotted back in.

SHIFT pipeline: Inspect, Policy, Transform, Reconstruct

1. Inspect

Detect every image in the payload. Extract format via magic bytes, read dimensions, measure file size. Handles base64 data URIs, raw base64, and URL references.

2. Evaluate

Load the provider profile. Compare each image against constraints. Apply drive mode rules to determine actions: resize, recompress, convert, rasterize, or drop.

3. Transform

Execute actions with Lanczos3 filtering. Rasterize SVGs via resvg. Convert BMP/TIFF to PNG. Recompress bloated images to meet size limits.

4. Reconstruct

Rebuild the payload with optimized images slotted back into their original positions. Output a valid, provider-safe JSON request ready to send.

Token savings

Real numbers from the two major multimodal providers.

Scenario	Before	After	OpenAI tokens	Anthropic tokens
4000x3000 hero (balanced)	4000x3000	2048x1536	765 → 765	1,568 → 1,568
4000x3000 hero (economy)	4000x3000	1024x768	765 → 765	1,568 → 1,082 -31%
1254x1254 icon (economy)	1254x1254	1024x1024	765 → 765	1,568 → 1,430 -9%
SVG diagram → rasterized PNG	SVG	512x256 PNG	255 → 255	0 → 198 enabled

Token estimates based on published provider formulas. OpenAI uses tile-based counting (512x512 tiles); Anthropic uses pixel-based (w×h/750, 1568px cap).

Three drive modes

Choose your optimization level per request.

PERFORMANCE

Maximum fidelity

Minimal transforms. Only enforce hard provider limits (max dimension, max file size). Preserve original quality.

BALANCED · DEFAULT

Smart optimization

Moderate optimization. Resize oversized images, recompress bloated files. Remove obvious waste while keeping quality.

ECONOMY

Maximum savings

Aggressive optimization. Downscale everything to 1024px, drop excess images, minimize token usage at all costs.

Works with any AI agent

Plugin, proxy, middleware, or CLI — pick the integration that fits your stack.

OPENCODE PLUGIN · RECOMMENDED

Zero-config for OpenCode

Auto-starts the proxy on every OpenCode launch. No manual setup per session.

// opencode.json
"plugin": ["@shift-preflight/opencode-plugin"]

PROXY

Universal agent proxy

Transparent HTTP proxy for Claude Code, Codex CLI, Gemini CLI, Aider — any agent that accepts a base URL.

ANTHROPIC_BASE_URL=http://localhost:8787

MIDDLEWARE

AI SDK drop-in

In-process middleware for any Vercel AI SDK app. Zero config, auto-detects provider.

shiftMiddleware({ mode: "balanced" })

CLI

Pipe & transform

Pipe any JSON payload through shift-ai directly. Works with any language or toolchain.

cat request.json | shift-ai -p anthropic

OpenCode plugin: @shift-preflight/opencode-plugin · Runtime package: @shift-preflight/runtime · Setup guide →

Install

Single Rust binary, zero dependencies. Pick your method.

Homebrew

Recommended for macOS/Linux

brew tap alohaninja/shift && brew install shift-ai

Quick install

curl one-liner

curl -fsSL https://raw.githubusercontent.com/alohaninja/shift/main/install.sh | sh

Cargo

From crates.io

cargo install shift-preflight-cli

AI Agent Skill

Teach your AI coding agent to use SHIFT automatically

npx skills add alohaninja/shift

Pre-built binaries → · Full guide →

Stop wasting tokens
on oversized images

One binary, three integration modes. Works with OpenAI, Anthropic, and Google out of the box.

Install SHIFT View source Read the guide

Multimodal preflightfor AI APIs. Automatic.