Capture One AI Windows
Summary: This MCP server lets a vision-capable AI inspect and natively edit Capture One through structured tools, while managing selections, previews, layers/masks, and operation evidence.
Inspect app context, target, capabilities, shared workflow, and unresolved operations.
Read the current photo/target, native parameters, module guidance/units, and observation tokens.
Apply qualified edits for exposure/contrast/brightness/saturation/white balance/HDR, curves, levels, color mixer, color balance, layers, and masks.
Manage masks/layers: create subject/background masks, invert, clear, fill, copy/combine, feather, linear/radial gradients, luma range, and select submask.
List/select images in the active collection, navigate next/previous, create candidate copies, and prepare tools/adjustments; edits are restricted to registered copies.
Render native or mask previews asynchronously, use recent PNG cache, and map coordinates between preview and RAW.
Inspect operation logs or reconcile current evidence without dispatching recovery writes.
Supports the OpenAI Responses API for direct AI photo-editing workflows: users provide an API endpoint and key, discover available models and reasoning efforts, and send photo previews and statistics to analyze images or make edits on protected Capture One variant copies.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@Capture One AI Windows Check the current photo, raise exposure a bit, and show me a preview."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Capture One AI · Windows / macOS
0.4.0a4 收录 Mac 原生桥和波形图会话同步。 MCP 和波形图自动跟随当前 Session,原生 PNG 与分析共用经过 ICC 转换的 sRGB 像素。升级时请同步重建 Mac 原生桥。 下载源码包 · Mac 安装 · 实测记录
0.4.0a3:Windows MCP/CLI 可选择副本或原版本。 默认保留副本流程;用户明确授权直接编辑时,images(prepare_edit, edit_mode="original") 登记当前版本,不创建副本、不重置照片。每张照片仍单独绑定 Session、变体与新观察凭据。API 修图窗口和 Mac 仍使用副本流程。见 接口说明。
Windows 和 Mac 的原生预览与波形取样均已在指定宿主版本上验证 sRGB 转换,PNG 内嵌 ICC。 色彩验证与升级说明
0.4.0a1:合并原生 Mac 开发预览与现有 Windows 客户端。 Windows 保留 16.8.6.3784 的现有功能;Mac 在 Apple Silicon / macOS 15.7.4 / C1 16.8.6.25 / Session 完成代表性实测,包含原生调色、图层/蒙版、RGB 波形和菜单栏常驻入口。Mac 是版本绑定的源码预览,尚非独立签名安装包;具体缺口见 Mac 当前范围。
安装时让桌面 AI 先询问是否创建桌面波形图入口。 同意后才创建,不同平台共用同一工具接口。复制安装提示 · Mac 安装 · Windows 安装。
直接填 API 对话修图,也可以接入桌面 AI 的 MCP;配套独立波形图和蒙版肤色分析。 让视觉 AI 通过平台原生接口读取和调整 Capture One,并取得 C1 实际渲染的预览。提供 API 修图窗口 + MCP stdio + JSON CLI,共用一套原生执行接口。
0.3.0a3 新增 API 修图窗口: 填写 API 地址和密钥 → 从服务读取模型 → 查询思考等级 → 输入需求,选择“仅分析”或“在副本修图”。无需另外安装聊天客户端。支持 OpenAI Responses API 及实现相同图片/工具调用协议的服务;仅有 Chat Completions 的接口暂不支持。模型和等级不使用固定清单,服务未提供的能力明确标为未知。
0.3.0a3 预览版,实测 Capture One 16.8.6.3784 / Windows / Session。 本机 C1 副本读写、预览和蒙版分析已实测;新 API 客户端的真实云端模型联调尚未完成,不能据此保证所有模型或第三方兼容服务可用。使用 C1 内部托管对象和原生处理内核,不是官方公开稳定 SDK。其他 C1 版本、Catalog 和多照片批量任务尚未完成兼容验证;启动器拒绝未验证版本。
Windows 自动匹配桥与 Session 自动绑定:本机安装、编译和离线检查已完成,待运行切换与实机验证。 新入口发现本机 C1,按兼容清单准备完整版本包,统一 MCP/CLI/AI/波形图所用版本;默认跟随当前原生 Session 的路径和 UUID。保留最新 sRGB 取图修复。检查范围和运行切换条件见 便携桥说明。
Native Capture One editing for Windows with an optional direct API chat window, MCP and a JSON CLI. Enter a Responses API key, discover models and reasoning efforts from your provider, and edit protected variants using the existing native bridge. Includes scopes, mask-aware analysis and ROI statistics. Cloud model integration is pending live validation. Unofficial, version-specific, MIT-licensed.
选择使用方式
方式 | 怎么开始 |
直接填写 API | 安装时加 |
接入已有桌面 AI | 将生成的 |
只看波形/直方图/矢量图 | 安装时加 |
API 修图会向所填服务发送照片预览、统计和必要的文件标识,不上传 RAW 原文件,并消耗该服务的 API 额度。密钥默认仅本次使用,也可选择 Windows 本机加密保存。只在登记副本上修改,保留原版本。
Related MCP server: Photoshop-MCP-IMT
功能
API 对话修图:动态模型与思考等级查询、只读分析/副本编辑、停止后续操作、用量显示;原 MCP 接入方式保持兼容。
曝光、对比度、亮度、饱和度、白平衡、HDR。
RGB / R / G / B 色阶;RGB / R / G / B / Luma 五通道原生曲线。
Basic、Advanced、Skin 混色器;主色轮和阴影/中间调/高光色彩平衡。
调整图层创建、选择、透明度、改名、开关、复制和排序。
主体/背景 AI 蒙版;反相、复制、组合、羽化、渐变和亮度范围。
模块重置与快照复用;当前集合内按 UUID 选图和翻页。
C1 原生照片/灰度蒙版预览,长边 128~1600 像素;预览与 RAW 坐标映射。
独立波形窗口和
analyze_image:亮度波形、RGB Parade、线条直方图、带肤色参考箭头的矢量图,以及供 AI 读取的同源数值、ROI 和参考差值。见 启动与接口说明 和 验证记录。0.3.0a1 新增;旧 0.2.0a1 ZIP 保持原样。0.3.0a2:跟随当前图层蒙版分析+预览中拖动框选。 选背景层看整图,选调整层只统计其蒙版覆盖的最终画面;波形、直方图、矢量图和 AI 数值使用同一覆盖权重,保留羽化。
只看皮肤区域怎么用
在 C1 选中覆盖皮肤的蒙版图层,在分析窗口切到“矢量图”。窗口默认开启“⋯ → 跟随当前图层蒙版”。
要进一步缩小范围,选择“⋯ → 在预览中框选区域…”,拖动矩形后点击“应用”。统计范围为矩形与蒙版的交集,不修改 C1 裁剪或蒙版。
选回 Background 即分析背景层范围;若之前设置了矩形框选,可在框选窗口点“恢复整个画面”并应用。取消“跟随当前图层蒙版”可忽略蒙版。
肤色参考线不会自动找脸或识别皮肤。主体蒙版可能包含衣服,分析纯肤色时需要相应的皮肤蒙版。空蒙版显示没有覆盖像素,绝不自动退回整图。
AI 分析当前图层时调用 analyze_image(analysis_scope="active_layer", include_vectorscope=True);省略 analysis_scope 保持旧整图语义。获取窗口同一份快照时传 analysis_id,详见 AI 安装与使用提示。
全局和局部调整共用同一接口。正常流程不导入预设、不截图识色、不发送键鼠输入,不要求展开面板。清晰度、锐化、降噪、镜头/几何及人像精修不在当前范围;默认不切文档、不导出。
安装与启动
希望让桌面 AI 帮你安装、把波形图启动器放到桌面并与 C1 一起打开,可复制 AI 安装提示。人用窗口,AI 用 MCP;两者共用同一图像分析核心,窗口关闭不影响 AI 读取数值。
Mac 使用 Apple Silicon Python 3.11+:python3.12 scripts/install_macos.py。同意创建波形桌面入口后加 --desktop-scopes;原生桥本机构建与授权步骤见 Mac 安装。以下保留 Windows 安装步骤。
需要 64 位 Windows、有效授权的 C1 16.8.6.3784、64 位 Python 3.11+、PowerShell 7.6+ 与 .NET 10 Windows Desktop Runtime。本机验收 Python 3.14.7 / PowerShell 7.6.5,其他 Python 版本未逐项实测。安装依赖需要联网,运行库和 C1 均不随包分发。
下载 Releases 的 Windows ZIP,解压到固定目录。在该目录的 PowerShell 7 中执行:
.\Install.ps1 -WithAI -WithScopes -SessionPath 'D:\Photos\MySession\MySession.cosessiondb'
.\Start-CaptureOne.ps1
.\AI.ps1已有项目和 Session 配置时,执行 Install.ps1 -EnvironmentOnly -WithAI -WithScopes 补装 API 窗口。会增加“Capture One AI 修图”桌面入口;打开后填入自己的 API 配置,建议先点“仅分析”。不使用内置 API 窗口时可省略 -WithAI。
需要配套窗口和桌面入口时,安装命令加 -WithScopes。已配置项目可执行 Install.ps1 -WithScopes -EnvironmentOnly;它会创建 Capture One + AI、Capture One 图像分析 两个快捷方式。前者启动 C1 并按设置联动窗口,后者可随时单独重开窗口。窗口“⋯”菜单可关闭自动联动;原厂快捷方式保持原样。详见 启动模式。
在 C1 打开配置的 Session,选中照片。首次加载桥接时应先正常退出 C1,再用本项目启动器启动;仅打开普通 C1 快捷方式不会加载桥接。启动器默认不会关闭已运行的 C1;确需正常重启时显式传 -Restart。见 安装说明。
使用外部桌面 AI 时,将安装器生成的 config/mcp-server.local.json 加入该客户端;内置 API 窗口无需另外导入 MCP 配置。核心字段是本机 Python 路径和 -m c1ai mcp。本版提供原六个工具及新增 analyze_image;旧 0.2.0a1 Release 为六个。外部客户端须支持图片块并使用视觉模型。见 AI 接入。
一条编辑通道
API 修图窗口 / 外部视觉 AI → 同一 MCP 工具 / CLI → 串行执行器与操作记录 → C1 内部命名管道
→ C1 原生参数/图层/蒙版对象 → 参数回读 + C1 原生预览context 确认环境;read(fields=[]) 获取当前照片;preview(kind=native) 实际看图。
images(create_candidate) 创建并登记变体副本;Windows 用户明确授权时可用 images(prepare_edit, edit_mode="original") 登记当前版本直接编辑。
按画面选择模块:read → adjust → preview,使用读取返回的目标和观察凭据。
全局完成后再决定是否加局部蒙版;达到目标即停止。
参数使用绝对值,同模块相关参数合并提交。每次写入有唯一编号;超时先查询原编号,不重复派发。只有登记候选可写,Windows 可登记副本或明确授权的当前原版本;参数回读不等于视觉效果合格。
包内文档
继续开发请克隆本仓库最新源码,先读 AGENTS。Windows 与 Mac 保留独立宿主实现,工具和分析核心共享。Mac 已实现与实机验证范围、复杂蒙版限制及版本绑定见 MAC_STATUS,安装入口见 INSTALL_MAC。本次发布为双平台源码包;旧 Windows ZIP 不被替换。
文档 | 内容 |
安装、启动、配置、升级和排错 | |
通用 MCP / CLI 接入与 AI 调色流程 | |
直接填写 API、动态模型/思考等级、对话修图和验证边界 | |
六个工具、请求示例、单位和返回值 | |
混色器、色阶、图层、蒙版和模块复用 | |
选图、原生预览和坐标映射 | |
已验证范围与未完成事项 | |
原生调用与原图保护 | |
开发和发布包生成 | |
开发 AI 接手顺序、原生优先和原图保护原则 | |
Mac 迁移入口、可复用代码、已知缺口与验收顺序 | |
Mac 原生开发入口、实机结果与剩余限制 |
运行状态在当前用户 %LOCALAPPDATA%\CaptureOneAI,不进入仓库。分发包不含照片、预览、Session、私人配置、连接令牌或历史操作数据库。Windows 原生预览只保留最近 8 张缓存,不是永久成片导出。
MIT License · Third-party notices
Mac 快速窗口直接读取 C1 当前完整裁切的 RGB 缓冲,持续更新代表值约 0.14~0.16s;精确局部分析仍按需计算,尚不承诺所有请求低于 0.5s。校样、位深及其他边界见 图像分析。
Available Tools
6 toolsadjustA
Apply one qualified numeric or structured module using its exact observed module token.
Structured modules: curves {channel,points:[[x,y],...]}; color_mixer {mode:'basic',color,hue?,saturation?,lightness?} or {mode:'advanced'|'skin',action:'add'|'update'|'remove',rgb? ,index?,hue?,saturation?,lightness?,smoothness?,uniformity_hue?,uniformity_saturation?,uniformity_lightness?,enabled?,inverted?}; levels {channel,input_black?,input_white?,midtone?,output_black?,output_white?}; color_balance {wheel,x?,y?,lightness?}; layers {action,layer_ref?,creation_operation_id?,opacity?}; masks {action:'create_subject'|'create_background'}. Layers also support rename {layer_ref,name}, set_enabled {layer_ref,enabled}, duplicate {layer_ref,name?}, move {layer_ref,index}; index is zero-based excluding Background. Masks on the selected adjustment layer: invert, clear, fill {density:0..1}, copy {source_layer} replaces the entire mask; combine {source_layer,mode:add|subtract|intersect}; feather {radius:0..300}; linear_gradient {start:[x,y],stop:[x,y],midpoint:0.01..0.99}; radial_gradient {center:[x,y],radius_x,radius_y,falloff:0..1,angle:-360..360,inverted:bool}; luma_range {shadow,shadow_falloff,highlight,highlight_falloff,radius,sensitivity,inverted}. Gradient coordinates/radii use normalized RAW dimensions, not rotated viewer coordinates. Luma endpoints 0..255 ordered shadow_falloff<=shadow<=highlight<=highlight_falloff, radius 0..300, sensitivity 0..100. select_submask {index} uses the native indexes returned by masks read; -1 selects the whole mask. Brush/gradient operations follow C1's selected submask semantics. Luma range replaces existing luma components, preserving the other masks. Each color module also accepts {action:'reset'} (C1 defaults) or {action:'paste',source_observation_token} (a saved complete read of the same module). Read the destination module before applying; source snapshot may belong to another layer/photo. Native qualification remains required. No presets or style files are created. Scalar parameters, levels, all mixers, color wheels and curves share the C1 internal model endpoint for Background and the selected adjustment layer. Select a layer then re-read the module; target.layer binds every write to that layer. Native AI masks use AutoMaskingManager without requiring the tool panel to be visible. Levels endpoints use 0..255, midtone -1..1. Advanced H/S/L limits are +/-30, -100..200, +/-100; Skin +/-5, +/-30, +/-10. Smoothness 1..30; Skin uniformity 0..100. Advanced/Skin may also update hue_low/high and saturation_low/high: signed offsets from the source color, low<=-2, high>=2, hue span<=359 degrees. Skin has one range. Read gives current native ranges and indexes; no screenshot or UI color picking is required.
Inputs: patch (required), module (required), target (required), observation_token (required), +1 more. Returns: object
contextC
Read application, target, capabilities, shared editing workflow and unresolved operations.
Inputs: no named parameters. Returns: object
imagesC
List the active native collection (including offscreen photos), or select by variant_uuid.
Follow next_offset with the same collection_token. next/previous follow the filtered collection. create_candidate clones the selected variant; edits remain restricted to registered copies.
Inputs: action (required), tool, limit, scope, +7 more. Returns: object
operationsC
Read full logs or reconcile current evidence without dispatching recovery writes.
Inputs: action (required), request_id, operation_id. Returns: object
previewB
Default native renders current C1 adjustments asynchronously, without export or focus.
kind=mask renders the selected layer's whole composite mask. max_size is 128..1600. map_points maps {x,y} pixel-edge coordinates using C1's nonlinear transform; use the same kind/max_size and returned preview_id. Divide RAW x/y by raw_dimensions to make normalized gradient coordinates. A changed photo, layer or revision invalidates the id. Only the most recent preview job is retained; eight PNG cache files are retained. Legacy kind=render exports, kind=viewer/action=capture samples the display.
Inputs: target (required), observation_token (required), kind, action, +5 more.
readA
Read native parameters with provenance, module guidance/units and an observation token.
Read fields=[] for target/preview without loading adjustment parameters. Read module=... before edits. Parameter targets are absolute, not deltas; skip modules without a visible purpose.
Inputs: fields, module, target, modules. Returns: object
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 6 tools
The broad purposes are mostly separable: adjust writes, images manages collections, preview renders, and read returns parameters. However, context, read, and operations all expose inspection/state information, and context explicitly includes unresolved operations while operations handles log/reconciliation evidence, creating boundary ambiguity.
Names mix imperatives and nouns: read and adjust are verbs, while context, images, preview, and operations are noun-like or ambiguous. There is no predictable verb_noun pattern, making the set feel less systematic even though each name is individually readable.
Six tools is a reasonable size for a desktop-editing integration, with each tool covering a major slice of the workflow. The small surface is appropriate given that adjust carries most of the editing complexity.
The surface covers context/state, parameter reads, adjustments, image selection, previews, and logs, which is substantial. But there are notable lifecycle gaps: no explicit export/save tool, no undo or recovery-dispatch operation, and limited image-management actions beyond selection and cloning.
Maintenance
Related MCP Connectors
AI-cull, edit and export RAW photos locally with the imagic desktop app, from any MCP client.
Hosted MCP server for AI photo editing: nothing to install, no desktop photo editor needed.
Compose, edit and render video with your AI agent. Turn footage into finished edits with content-aware reframing, word-timed captions and motion graphics. Hosted MCP with OAuth sign-in; cloud processing uses credits.
Flickr MCP server for ChatGPT, Claude and any AI agent: list albums, upload photo, edit title and tags, move photos between albums, change privacy, reply to comment and add photo to group in your own Flickr account. Cotlet works in the account through your Cloud Browser, a real Chrome where you sign in to Flickr yourself; Cotlet does not store your password and you need no Flickr API key. About the official Flickr API: Flickr has a full REST API with methods for photos, albums, comments, groups and uploads, but no official MCP server (checked 17 Oct 2026). Only Flickr Pro subscribers can request new API keys, and commercial use needs a separate commercial key, by prior arrangement with Flickr. Setup and FAQ: https://cotlet.dev/integrations/flickr-api-mcp-server
Related MCP Servers
- AlicenseAqualityCmaintenanceAn MCP server that enables automating Capture One Pro on macOS, providing tools for reading and editing images, managing layers/masks, recipes, and processing with risk-tiered write gating.184MIT
- AlicenseBqualityBmaintenanceLets AI agents drive a running Adobe Photoshop: 135 tools behind a 10-tool facade, a UXP bridge that keeps working while Photoshop shows a dialog. Works with every UI language Adobe ships and checks its work by sight. A large skills library, built for Adobe's generative features: Generative Fill, Expand, Upscale, Firefly and partner models.41351,000 npm6Apache 2.0
- AlicenseNot gradedqualityBmaintenanceEnables AI agents to run cross-platform batch photo-processing workflows—including RAW decoding, adjustments, noise reduction, white balance, LOG restoration, EXIF tagging, deduplication, selection, HTML galleries, folder watching, and presets—through a unified versioned JSON contract. It also exposes MCP tools for processing, suggestions, autopilot, indexing, search, selection, HDR, face blurring, and more.MIT
- AlicenseNot gradedqualityCmaintenanceEnables AI agents to drive Adobe Lightroom Classic directly — searching and organizing catalogs, reading and writing the full set of develop parameters, building geometric and AI-generated masks (sky, subject, mountains, water, people), applying presets, and rendering previews, histograms, and exports. All capabilities are exposed as 167 MCP tools over a Lua plugin bridge, with write-confirmation sentinels so silently dropped edits surface as explicit errors.Apache 2.0