AfterEffects-MCP-IMT + SKILLS — After Effects for AI agents

io.github.Immersive-Media-Technologies由 GitHub 帳號 Immersive-Media-Technologies 發布v1.1.1更新於 Nov 7, 2026

After Effects for AI agents: every UI language, vision for mattes and timing, skills, generative AI

已驗證STDIO僅桌面開發者工具媒體與設計

概覽

AI 產生的概覽

讓 AI 助理讀取並編輯正在執行的 Adobe After Effects 專案——合成、圖層、效果、關鍵影格、遮罩、算圖——並能查看算出的畫面。

可以做什麼
  • 探索並執行 After Effects 操作(約 270 項,分為 24 類),可單次或批次執行
  • 讀取專案、合成、圖層與版本資訊,以及合成快照與差異比對
  • 將單一影格算成 PNG,並與先前的影格比較
  • 在 macOS 上依畫面繪製物件遮罩筆畫,並逐格將 AI 生成圖層與實拍素材對齊
  • 儲存專案,並透過 JSON 完整匯入匯出專案
  • 檢視並關閉阻擋腳本的模態對話框
適用情境
適合用自然語言讓助理檢查、建立或修正真實的 After Effects 工作:找出遺失素材或表達式錯誤、加入效果與關鍵影格、製作標題動畫、測試外掛、算圖檢查影格,尤其適用於非英文介面的 After Effects。
執行需求
同一台 macOS 或 Windows 電腦上執行 After Effects 2024–2026;除使用 Claude Desktop 擴充功能外需 Node.js 22+。macOS 需為 AI 用戶端授予「自動化」權限,處理對話框還需「輔助使用」權限;After Effects 需允許指令碼寫入檔案並存取網路。選用變數:AE_MCP_EXE、AE_MCP_READONLY、AE_MCP_PROJECT_LOCK、AE_MCP_CHECKPOINT_DIR、AE_MCP_ENABLE_EVAL、AE_MCP_ALLOW_CATEGORIES、AE_MCP_RUNTIME_DIR、MCP_TIMEOUT。
安裝前請注意
它會修改真實專案,並把專案內容(名稱、表達式、素材路徑、算圖預覽)傳送給你使用的 AI 服務,請在副本上操作,並就機密工作確認服務商的資料政策。除非需要原始指令碼,否則保持 AE_MCP_ENABLE_EVAL 關閉;eval.run、專案新建/開啟/還原、清除與合併、偏好設定和 render.start 無法用 Cmd+Z 復原。在 macOS 上它會以真實滑鼠事件與視窗擷取來操作 After Effects。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 AfterEffects-MCP-IMT + SKILLS — After Effects for AI agents,將其新增到工作區。
  2. 為需要使用其工具的對話啟用該服務。

Desktop only,透過 STDIO。 STDIO 服務會啟動本機處理程序,因此需要 SourceWeft 桌面主機。

其他 MCP 客戶端

參照 儲存庫 中的啟動說明。

README

[AfterEffects-MCP-IMT + SKILLS — After Effects for AI agents]

AfterEffects-MCP-IMT + SKILLS — After Effects for AI agents, built to be safe to hand over

Let an AI assistant (Claude Desktop, Cursor, ChatGPT, Google Antigravity or any MCP client) work inside your running Adobe After Effects: read the project, build and edit comps, layers, effects, keyframes, text, masks and expressions, render check frames — from plain language.

Every Adobe UI language · vision: mattes and time by sight · a large skills library · built for AI and generative features. AfterEffects-MCP-IMT works with every UI language Adobe ships, and the agent sees the frame: it renders frames and looks at them, draws Object Matte on what it sees — a lasso around each person to keep, ⊕ / ⊖ points for what the matte missed or caught — reads the strokes back and checks a preview that marks the gaps and every person it was told to keep (macOS); it syncs an AI-generated layer with the plate frame by frame — finds the matching frame for every frame, then cuts and shifts the layer so each lands on it — and checks every frame. After Effects' scripting has no API for Object Matte strokes, none of the other After Effects MCP servers we compared draws them, and Adobe's own AI Assistant for After Effects (beta) sees a single frame and writes scripts it does not run. It comes with a large skills library and is built for Adobe's AI and generative features — Roto Brush and Object Matte, Content-Aware Fill, generated image sources animated as native layers.

Two parts, made to work together:

  • The MCP server — ~270 small, safe operations behind two tools, with a dry run, one undo group per call, checkpoints, a project lock and honest errors; Object Matte by sight on macOS (ae_object_matte_ui drives After Effects' own Object Matte tool — there is no API for its strokes).
  • The skill (skills/ae) — what the agent should know: the order of work, what a script cannot do in After Effects and how to get around it safely, step-by-step recipes (camera tracking, 3D content on footage, plug-in tests, render, motion graphics, masks), and plug-in knowledge files.

Built for localized After Effects, not just English. Most MCP servers and agent skills for Adobe apps assume the English UI — on a German, Japanese or Russian install their effect, template and menu names stop matching. AfterEffects-MCP-IMT addresses After Effects by its own identifiers (matchNames, command ids) and translates English template names through After Effects' own dictionaries, so the same agent and the same skill work on every UI language After Effects ships — verified live.

By Immersive Media Technologies.

macOS and Windows · After Effects 2024–2026 (26.5 verified on both, every UI language) · Node.js 22+ (none needed with the Claude Desktop extension — verified on macOS and on the Windows Store build of Claude Desktop).

[!CAUTION] This edits real After Effects projects and sends project contents (names, expressions, footage paths, rendered previews) to the AI service you use. Work on a copy, check your provider's data policy for NDA work, and keep AE_MCP_ENABLE_EVAL off unless you need raw scripting.

What you can ask for

  • "Point out anything broken in this project — missing footage, fonts, expression errors."
  • "Add a glow to every text layer in Intro, 30 % lighter than now, and show me frame 2.0."
  • "Import this Illustrator file and animate the title with the text-pop preset."
  • "Track the camera on this clip, put a 3D null on the floor and parent the logo to it." (skill)
  • "Here is a reference frame: build this lower third natively, with a Speed slider and the texts in Essential Graphics." (skill)
  • "Test this plug-in on a solid, dump its parameters, render a frame and tell me if it works." (skill)

The agent combines the operations itself and checks the project between steps; the skill keeps it on the rails.

Install

You need After Effects running on this computer and two permissions:

  • macOS — Automation: the first call asks whether your AI client may control After Effects (a refusal shows up later as error −1743: System Settings → Privacy & Security → Automation).
  • Scripting file access: After Effects → Settings → Scripting & Expressions → Allow Scripts to Write Files and Access Network.
ClientHow
[圖片] Claude Desktopdownload aftereffects-mcp-imt.mcpb, double-click, Install (Windows: drag it onto Settings → Extensions). No Node needed.
[圖片] CursorAdd to Cursor — one click: Cursor opens and writes the config (or put the snippet below into ~/.cursor/mcp.json).
[圖片] Google AntigravitySettings → Customizations → Installed MCP Servers → View raw config → add the snippet below.
[圖片] ChatGPT (remote servers only)on the After Effects machine: npx -p @immersive-media-technologies/aftereffects-mcp-imt aftereffects-mcp-imt-chatgpt — it starts the server behind a tunnel and prints the URL to paste into ChatGPT.
Any MCP clientthe snippet below in its mcpServers.
json
{  "mcpServers": {    "aftereffects": {      "command": "npx",      "args": ["-y", "@immersive-media-technologies/aftereffects-mcp-imt"],      "env": { "MCP_TIMEOUT": "120000" }    }  }}

From source, with a live self-test (the route we use):

bash
git clone https://github.com/Immersive-Media-Technologies/aftereffects-mcp-imt.gitcd aftereffects-mcp-imt && ./install.sh        # Windows: .\install.ps1

The installer checks After Effects and Node, builds, pings the running AE, names what is wrong (−1743, scripting access, Node path) and — after asking — writes the entry into Claude Desktop's, Cursor's and Antigravity's configs, backing them up.

If something does not connect

  • spawn npx ENOENT on macOS: GUI apps do not see your shell PATH — put the absolute path of a Node from Homebrew or nodejs.org in place of npx (nvm's Node is invisible to them).
  • Set MCP_TIMEOUT to 120000: a cold After Effects start does not fit the default 30 s.
  • A timeout with a DIALOG_OPEN message: After Effects is showing a dialog (save prompt, missing footage, a plug-in warning) — the error quotes it; the agent can close it with ae_dialogs or you click it away. On macOS reading and closing dialogs needs Accessibility permission for the client app.
  • Windows: AE_MCP_EXE must point to AfterFX.exe when After Effects is not under C:\Program Files\Adobe. No supported GPU (a VM): dismiss the «System Compatibility Report» first. The .mcpb needs no Node.js on the machine — Claude Desktop runs it on its built-in Node (verified with the Microsoft Store build on Windows 11 ARM). Server log: %LOCALAPPDATA%\Claude\logs\mcp-server-<name>.log; mailbox: %TEMP%\mcp-aftereffects.
  • If your client caps tool output, raise it to 50 000 — ae_project_info on a real project is bigger than the usual 25 000.

How it stays safe

Dry runae_do {dryRun: true} shows the ExtendScript that would run and never touches After Effects.
One undo group per callevery ae_do — even a 30-step batch.run — is one Cmd+Z. batch.run {atomic: true} undoes itself when a step fails and says so (rollback.verified).
Checkpointsproject.checkpoint keeps a copy of the .aep next to the project; project.revert brings it back after a crash, a relaunch or a bad idea — things undo cannot do.
Project lockAE_MCP_PROJECT_LOCK=/path/to/one.aep — the agent may change only that project; your other open work is untouchable.
Read-only modeAE_MCP_READONLY=1 for a first session: inspection and check frames only.
Honest errorsevery failure has a code (DIALOG_OPEN, TIMEOUT, INVALID_ARGS, OPERATION_FAILED …) and a hint; a batch reports which step failed.
Nothing installed inside AEthe server talks to After Effects from outside (osascript / AfterFX.exe -r); uninstall = delete the folder.

What to forbid on the client side, if you gate at all: what Cmd+Z cannot revert — eval.run, project.new / open / revert, the purge/consolidate operations, pref.*, render.start. The list and a reference implementation are in docs/AGENT-POLICY.md.

How the agent sees

Adobe gives outside agents no picture: its scripting returns a project's structure, not its pixels; its own connector for AI assistants (Adobe for creativity) does not cover After Effects; and the After Effects AI Assistant (beta) sees «a single screenshot of a single frame» and does not run the scripts it writes.

AfterEffects-MCP-IMT gives the agent two kinds of sight. The model's: every frame it needs comes back as an image — ae_render_frame for the comp, a solo render of one layer for a matte — and the model decides from the picture what to do and where. The server's: where After Effects has no scripting API (Object Matte strokes are mouse strokes in the Layer panel, nothing else), the server looks at After Effects itself — it captures the window, finds the picture in the Layer panel, maps the model's picture coordinates onto the screen and acts with real mouse events, then looks again to confirm.

Object Matte by sight (ae_object_matte_ui, macOS, AE 26):

  1. The layer is rendered solo at the chosen time; the model answers in picture coordinates — a closed lasso around each person or object to keep, ⊕ points for what the matte missed, ⊖ for what it caught by mistake.
  2. The server opens the layer, selects the Object Matte tool by its internal tool code (not by a hotkey or a localized name), draws lassos first and points after them, and finds the picture again after every stroke — After Effects shifts the view while it computes.
  3. The strokes are read back from the effect; a preview renders the matte over the frame, marks the holes inside the subject, and checks every person the model named at head, body and feet.
  4. The model fixes what the preview shows with more ⊕ / ⊖ strokes. For a matte that will cut an AI-generated layer, the edge is then pulled inside the silhouette (Shift Edge, Feather) by the effect's internal property names — After Effects shows two groups with the same labels.

Time by sight (recipes/layer-conform.md): an AI-generated layer drifts against the plate — repeated, dropped and frozen frames. Frames of both are rendered and compared one by one, the best match is chosen for every frame (scripts/conform.py refine-cuts, repeats and drops included), and the layer is cut into pieces and each piece shifted so every frame lands on its match — cuts only: a speed between two points would blend frames. Then every frame is checked again.

Everything else the agent reads as data:

  • ae_project_info / ae_comp_info / ae_layer_info — the full picture when needed.
  • comp.snapshot → edit → comp.diff — a cheap fingerprint before and after: which layers were added, removed or changed, and in what. Verification in a few dozen tokens.
  • ae_render_frame — a PNG the agent can look at; compareWith an earlier frame answers "did the picture change, and where" without opening the image.
  • property.sample — a camera's position / orientation / zoom at start, middle and end in one call; expression.validate and project.audit — errors before the user hits play.

The skill

skills/ae/SKILL.md is a short router the agent reads first; it points to the one file the task needs:

SKILL.mdthe call budget (what made a live plug-in test cost 76 calls an hour — and how not to), the order of work (know the state → checkpoint → operate → verify by values, then by frame), and a table of what a script cannot do in After Effects with the safe workaround for each.
recipes/camera-track.md3D Camera Tracker end to end: start the analysis from the menu, wait in one call, press «Create Camera» with a real click (a synthetic click crashed AE 26.5), set the ground plane, read the camera back as numbers.
recipes/attach-3d.mdput a splat, a model or a solid on tracked footage: null at the origin, parent or expressions, scale, and a script that measures slide in pixels instead of eyeballing it.
recipes/plugin-test.mdtest a third-party plug-in: inspect the files before installing, apply on a solid, dump parameters by matchName, feed it a model layer, render, time it, write the verdict.
recipes/render.mdframes, segment renders, templates on a localized AE, where files go.
recipes/motion-graphics.mdtitles, lower thirds, logo stings, infographics as an editable AE comp: breakdown table, CONTROLS rig, real easing, Essential Graphics, a frame per phase.
recipes/masks-roto.md, recipes/tracking-mocha.mdmasks and track mattes, Object Matte / Roto Brush, the Tracker panel and Mocha — what to script, what the user clicks.
recipes/layer-conform.mdsync an AI-generated layer with the plate before compositing: time drift → Time Remap keys, scale/shift → Transform per piece, colour → Levels, matte — measured on rendered frames by scripts/conform.py.
plugins/<name>.mdone knowledge file per plug-in (matchNames, how it takes input, what is UI-only, quirks, timing). First: irrealix Gaussian Splatting. Template in plugins/README.md.
scripts/only the UI steps no script can do: hardware-level mouse events, window and display geometry, menu clicks by localized name, dialogs, a one-call wait for an analysis, screenshots, a frame strip, the slide measurement. No script duplicates a server operation.

Use it as a skill in clients that support them (copy skills/ae into ~/.claude/skills/, or point your system prompt at SKILL.md). The motion-graphics recipe routes to topic references from five open MIT skill sets; they are fetched, not bundled — run skills/ae/fetch-vendor.sh once.

Tools

ToolPurpose
ae_catalog / ae_dodiscover the operations (24 categories) and run one — or several in one batch.run
ae_project_info / ae_comp_info / ae_layer_info / ae_version_info / ae_contextread
ae_render_frameone frame as PNG (viewer-accurate colour; compareWith for metrics)
ae_object_matte_uiObject Matte by sight (macOS): lassos and ⊕ / ⊖ points from a rendered frame, strokes read back, a preview of gaps and people
ae_save_project, ae_project_export_json / ae_project_import_jsonsave; whole-project JSON round trip
ae_dialogssee and dismiss the modal dialogs that block scripting

Full list with parameters: docs/TOOLS.md; the operations are listed by ae_catalog at runtime.

VariableDefaultMeaning
AE_MCP_EXEauto-detectthe After Effects .app bundle (macOS) or AfterFX.exe (Windows)
AE_MCP_READONLYoff1 → nothing may modify the project
AE_MCP_PROJECT_LOCKoffpath of the one .aep the agent may change
AE_MCP_CHECKPOINT_DIRnext to the projectwhere project.checkpoint keeps its copies
AE_MCP_ENABLE_EVALoff1 → eval.run (arbitrary ExtendScript) becomes available
AE_MCP_ALLOW_CATEGORIESallcomma-separated allow-list of operation categories
AE_MCP_RUNTIME_DIROS tempwhere the IPC mailbox lives
MCP_TIMEOUTclientset 120000 in the client config

Why this one

There are several After Effects MCP servers; most stop at "the AI can run ExtendScript". This one is built around what an autonomous agent needs — and what the person handing over a project needs: a small stable tool surface (two tools instead of 270), a real dry run, undo-safe operations, checkpoints and a project lock, honest errors, an installer that says what is wrong, and a skill that teaches the agent the parts of After Effects no API covers — on any UI language After Effects ships, where most servers stop at the English one. Verified on AE 2026, macOS and Windows.

How the After Effects agents compare on vision — as of 2026-11-06:

AfterEffects-MCP-IMTmcp-aftereffects (upstream)JUNKDOGE-JOE/after-effects-mcpEngine-Room-Games/after-effects-mcpAfter-Effects-mcp-proafter-effects-mcp-visionAdobe AI Assistant (beta)Filmit MCP (commercial)
Frames returned to the modelcomp frames and solo layer rendersyesyesyesyes, and window screenshotsyesa single screenshot of a single frameyes, when the model asks
Vision: acting on what the agent seesObject Matte drawn on what it sees — lassos, ⊕ / ⊖ points, strokes read back, a preview of gaps and people (macOS); a generated layer synced to the plate frame by framenononono — Roto Brush «only through screen control»mask paths from a separate SAM app, not Object Mattewrites scripts, does not run them; no Roto Brush or Object Matte in its helpmasks listed; no Roto Brush or Object Matte

Releases

Each version is described on the Releases page; every change is in CHANGELOG.md. Upstream fixes are merged from kumoproductions/mcp-aftereffects by topic.

  • v1.1.1 · 2026-11-06 — descriptions with vision next to the UI languages (Object Matte by sight, generated layers synced frame by frame), a dated vision comparison and a «How the agent sees» section.
  • v1.1.0 · 2026-11-06 — ae_object_matte_ui: Object Matte by sight (macOS) — lassos and points on a rendered frame, strokes read back, a preview image that marks gaps and the listed people; layer conform recipe (cut-only time, conform.py 1.2.0 with refine-cuts, the matte edge inside the silhouette).
  • v1.0.2 · 2026-10-28 — descriptions everywhere: every Adobe UI language, a large skills library, built for Adobe's generative features.
  • v1.0.1 · 2026-10-28 — product name AfterEffects-MCP-IMT (was AE-MCP-IMT); server name aftereffects-mcp-imt like the npm package; the MCP Registry entry also lists the npm package.
  • v1.0.0 · 2026-10-28 — every After Effects UI language (matchNames, command ids, template names through After Effects' dictionaries; verified live), corrected command list, Apache-2.0.
  • v0.5.5 · 2026-10-11 — the After Effects skill (skills/ae), checkpoints + revert, atomic batches, comp snapshot/diff, expression validation, project audit, dialog detection, project lock, frame compare.
  • v0.4.4 · 2026-10-04 — Windows: Claude Desktop from the Microsoft Store.
  • v0.4.3 · 2026-10-04 — Windows verified (AfterFX.exe -r launch path).
  • v0.4.2 · 2026-10-01 — ChatGPT gateway fix; v0.4.1 — ChatGPT in one command; v0.4.0 — one-click installs (.mcpb, Cursor, npm, installers).
  • v0.3.5 · 2026-09-30 — the AE Motion skill; v0.3.0 — first public release of the IMT fork.

A note for contributors: the offline transport tests (npm test) expect After Effects to be closed on macOS — with it open the live AE answers the stand-in requests and a few of them fail.

License

Apache License 2.0 — © Immersive Media Technologies.

  • Free to use, modify and distribute, commercial use included.
  • Attribution is required. Every redistribution of this software or of a work derived from it must carry the notice from NOTICE: AfterEffects-MCP-IMT by Immersive Media Technologies — https://github.com/Immersive-Media-Technologies.
  • No warranty.

Third-party notices (upstream MIT code): THIRD-PARTY-NOTICES.md.

來源:README.md,提交 80d37cb

工具

0
SourceWeft 尚未在沙箱中啟動過該軟體包,因此其工具未知。

版本歷史

2
  1. v1.1.1最新Nov 7, 2026
  2. v1.0.2Oct 29, 2026

評分

已滿足適用檢查項的 50%

該評分衡量此收錄條目的完整程度,依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價,也不表示伺服器安全或目前可用:可連通性是帶日期的檢查結果,評分也不包含安裝量或使用者評價。

  • 近 30 天內可連通25 分

    請保持端點能夠回應 MCP 握手;SourceWeft 每 30 天探測一次。

  • 已宣告認證方式10 分

    請在註冊中心條目中宣告憑證請求標頭,或保持端點能夠回應探測,以便判定其認證方式。

  • 已列出工具15 分

    請讓伺服器在無憑證時回應 tools/list,或發布可在沙箱中無憑證啟動的軟體包。

  • 描述必需15 分
  • 圖示10 分
  • 文件15 分
  • 授權條款10 分