Reads the photo, not just pixels
EXIF camera/lens/exposure data, orientation, GPS, dimensions, luminance histogram, clipping, white-balance cast, skin-pixel ratio and a Laplacian sharpness score.
Reads EXIF and image analysis, plans the edit with a local model, and executes it through MCP tool calls.
The assistant starts with the details a photographer would check: camera, lens, ISO, shutter, aperture, date, orientation, GPS, dimensions, histogram, clipping, white-balance cast, skin ratio and sharpness. It can also inspect a downscaled copy of the image.
It then plans an edit and executes it through the editor’s own MCP (Model Context Protocol) tool server. The fast rules planner works without a model download; an optional local LLM or vision model can plan open-ended requests. Every tool call is visible, validated and undoable.
A photo-aware local AI assistant. Reads EXIF, histogram, skin and sharpness data, then calls editor tools over MCP to adjust, crop, clean or export.
EXIF camera/lens/exposure data, orientation, GPS, dimensions, luminance histogram, clipping, white-balance cast, skin-pixel ratio and a Laplacian sharpness score.
Fast deterministic rules are the default. Optional Qwen2.5 0.5B or SmolVLM 256M backends can plan open-ended prompts; models load lazily and are cached after the first download.
The browser runs an MCP JSON-RPC server and client: initialize, tools/list, tools/call and structured results. Tool arguments are schema-validated before execution.
Tools add adjustment layers, apply recipes, auto-tone, crop, add a text watermark, remove an object, remove a background or export a JPEG/PNG.
The panel shows the photo analysis, the plan explanation and every MCP request/response. Nothing is applied silently.
Analysis and planning run locally. No photo is uploaded. Mutating tools use the editor’s history stack, so a bad plan can be undone.
No install step and no import queue. Open the editor, work on the document locally, and export the file your next tool expects.
Try it nowThe assistant reads the document and EXIF metadata, then builds a compact analysis summary for the planner.
Use a prompt like “brighten and warm” or a preset such as Clean distractions, Bright & airy, Reception flash cleanup or Export web JPEG.
Inspect each MCP tool call and its result, then keep the edit or undo it. Export tools produce a normal download.
These are the workflows the tool is designed around — not a list of theoretical features.
Straight answers about formats, privacy and limits.
No. Metadata parsing, pixel analysis and the MCP tools run in the tab. Optional local models download weights from Hugging Face once and then run through ONNX Runtime WebAssembly or WebGPU.
MCP is the JSON-RPC tool protocol between the local agent and the editor. The browser hosts a small MCP server exposing photo tools; the planner calls tools/list and tools/call, and the panel logs every exchange.
The default is a deterministic local rules planner with no download. You can opt into a small local text model (Qwen2.5 0.5B) or vision model (SmolVLM 256M) when you want open-ended planning. Model availability and speed depend on the device.
Tools are bounded and schema-validated. Mutating edits go through the editor store and undo history; exports create a download. There is no background or autonomous execution outside a run you start.
Open the editor and bring a real file. Your photo stays on your device.