Conduit for Photoshop MCP

The conduit between AI and Adobe Photoshop.

Built by Jag Journey, LLC.

What is it

Conduit gives any AI assistant (Claude, GPT, Grok, Gemini, Ollama, or any custom LLM) direct programmatic control over Adobe Photoshop. The AI opens documents, manipulates layers, runs Generative Fill, applies layer styles, exports for web, and saves the result. No copy-paste of Photoshop instructions. No "describe what you want and a human does it."

The AI also sees the canvas after every operation, so it can self-correct in the same turn.

Why it exists

An in-house Jag Journey workflow broke the existing community Photoshop MCP on May 9, 2026. Conduit is the replacement. Born from a real production session, hardened against the gaps the existing tools could not cover.

What you can do with it

Roughly 940 tools across 80+ Photoshop namespaces, all callable from natural language:

• Documents: open, create, save, save as, export in 12+ formats
• Layers and groups: create, find, duplicate, group, set blend modes, opacity, locks
• Transforms: scale, rotate, skew, perspective, warp, content-aware scale
• Selections, masks, channels, paths: the full Photoshop surface
• Layer styles: drop shadow, inner glow, stroke, gradient overlay, the works
• Color adjustments: curves, levels, hue/sat, color balance, vibrance
• Filters: every blur, sharpen, noise, stylize, render, distort and pixelate filter
• Firefly Generative AI: generative fill, expand, remove, object replace, text to image
• Neural Filters: skin smoothing, super zoom, colorize, landscape mixer, style transfer
• Web export: 2x/3x bundles, WebP, AVIF, JPEG XL, responsive srcset
• Print and color science: soft proofing, gamut warnings, ICC, LUT generation
• Camera Raw: Basic, Tone Curve, HSL, Split Toning, Color Grading, Detail, Lens
• Tethered shooting: drive supported cameras directly into Photoshop
• Variable fonts and OpenType deep features
• Accessibility: alt text, reading order, contrast audit, a11y-checked export
• Vision: describe image, detect faces, detect text, aesthetic score, dominant colors

Every tool returns structured JSON with real layer IDs, bounds, and error codes. Optional base64 PNG preview comes back in the same response so the AI's vision model can verify the result.

Featured workflows

Five end-to-end recipes the AI can drive today:

1. Jag Journey Client Case

"Open the hero background. Place a set of product photos into the layout frames, scale to fit, add a 4px white stroke and a 12px drop shadow on each. Duplicate the group, flip vertical, motion blur, gradient mask for the floor reflection. Use Generative Fill to remove a stray object. Save as PSD and export a 1024x1024 PNG."

2. Fix the doubled ceiling

"The top of the canvas has a doubled ceiling artifact. Select the affected band with a feathered rectangle, run Generative Remove with prompt 'clean continuous ceiling, no artifacts', screenshot the result, and if the artifact is still visible run a second variation."

3. Drop-shadow plus responsive web export

"Find the layer named 'product', add an outer drop shadow at 60% opacity, then export it at 1x, 2x, and 3x as PNG for web, plus WebP and AVIF at 2x."

4. Generative-Fill background swap on a product shot

"Select the sneaker with selection.subject, invert, run firefly.generative_replace_background with the prompt 'sunlit Joshua Tree desert at golden hour, shallow depth of field'. Show me the four variations side-by-side."

5. Soft-proof for print plus a11y-checked export

"Soft-proof the active document against US Sheetfed Coated v2, flag out-of-gamut areas. Then export as PDF with an accessibility report attached."

Compatible with every major LLM

| LLM | How | |---|---| | Claude Code / Claude Desktop | Native MCP support | | OpenAI Assistants / GPT | Via mcp-openai-bridge adapter | | xAI Grok | Via mcp-grok-bridge adapter | | Google Gemini | Via mcp-gemini-bridge adapter | | Ollama (free, local) | Via mcp-ollama-adapter | | Custom LLM | Use the published JSON tool schema |

The MCP server is LLM-agnostic. Build it once, every AI uses it.

How it works

You: "Place the Vogue cover into the back-wall acrylic frame, scale to fit, add a drop shadow"
         |
    Your AI client (Claude, GPT, Grok, Gemini, Ollama, custom)
         |  MCP protocol over stdio
    Conduit MCP Server (Node.js)
         |  WebSocket on 127.0.0.1:8765
    Conduit UXP Plugin (runs inside Photoshop)
         |  Photoshop UXP DOM API
    Adobe Photoshop 2022 or newer

Two key differentiators vs the existing community MCP:

1. Visual feedback every call. Every tool can return a base64 PNG of the canvas, so the AI sees the result and can self-correct.
2. Real return data, not undefined. Every tool returns structured JSON with layer IDs, bounds, errors. The existing MCP's execute_script returns no value; Conduit always returns the truth.

Status

Active development. Open beta. Get on the waitlist at jagjourney.ai.

Pricing and access

Conduit is built and maintained by Jag Journey, LLC. Pricing, licensing, sponsorship tiers, and enterprise plans live on the company site at jagjourney.ai.

The codebase is hosted privately on Jag Journey's self-managed git repository. This GitHub repository is the public landing page only; it does not contain source code.

Sponsor a feature

Conduit's v1.0.0 ship gate requires three named feature sponsors. Pick anything from the public roadmap, fund it, and your name lands in the CHANGELOG for the release that ships it. Direct invoice (USD wire/ACH) or Stripe Checkout. Reach us through the contact form: jagjourney.ai/contact/.

Get in touch

• Website: jagjourney.ai
• Contact form: jagjourney.ai/contact/
• Project owner: Jag Journey, LLC

License

Apache 2.0. Conduit's code (hosted privately) is open source; this README and any marketing copy here is also Apache 2.0 unless noted otherwise.

Built and maintained by Jag Journey, LLC.