Files
effectcraft/docs/control-protocol.md
Brandon Thomas 287c35dae7 Merge pull request #1109 from eliasx45/upstream/azerty-graph-zoom
# Conflicts:
#	crates/engine/src/commands/time.rs
#	crates/ui-egui/src/panels/timeline.rs
2026-10-13 04:29:47 -04:00

16 KiB

Control protocol

The desktop app exposes a control channel for agents, tests and the MCP bridge:

effectcraft --control 9877        # or EFFECTCRAFT_CONTROL_PORT=9877

The port must be a number from 1 to 65535, and --control overrides the variable. A bad port or an unknown option prints an error and the app exits with status 1 before it opens a window.

  • Transport: TCP on 127.0.0.1 only, one JSON object per line in each direction. Connections are persistent; requests on one connection are answered in order.
  • Request: {"id": 1, "method": "engine.execute", "params": {...}} (id is echoed back, params defaults to {}).
  • Reply: {"id": 1, "ok": true, "result": ...} or {"id": 1, "ok": false, "error": "message"}.
  • Requests run on the UI thread between frames. Input methods (ui.click, ui.key and similar) reply after the synthetic input has been processed. Methods that need an element that isn't drawn yet are retried for a few frames. A request times out after 60 s.
  • Only requests are read: every line must be a JSON object with a string method (blank lines are skipped). Anything else (text that isn't JSON, a JSON array or number, an object without method, invalid UTF-8, a line longer than 4 MiB) gets one error reply ending in "closing the connection", and the server closes the connection, so nothing sent after it runs. An HTTP request (for example a web page's cross-origin fetch to 127.0.0.1:<port>) therefore can't smuggle a command in its body: its request line is rejected first. At most 16 connections are served at once; further ones get an error line and are closed. The port has no authentication, so only enable it while you use it.

Quick test:

printf '{"id":1,"method":"engine.execute","params":{"command":"comp.new","params":{"name":"A"}}}\n' | nc 127.0.0.1 9877

Engine

Method Params Result
engine.execute {command, params?} The command's result. Every engine command (effectcraft-cli commands) runs exactly as it does headless, so it never opens a dialog: comp.new {} creates a default comp. UI-only ids (tool.*, view.*, window.*, playback.*, timeline.*, app.*) fall through to the menu dispatcher. Unknown top-level params are rejected, and the error lists the accepted keys.
ui.menu.invoke {id, params?} Like a menu click: commands with empty params may open their dialog (New Composition, Solid Settings, file pickers).
UI commands (via either) timeline.column {column, visible?} (av, keys, label, num, comment, switches, modes, parent, in, out, duration, stretch), timeline.sourceName {value?}, timeline.search {query}, timeline.reveal.<kind> / timeline.revealAdd.<kind> (the property shortcuts and Shift+shortcut: position, scale, rotation, opacity, anchor, effects, masks, feather, levels, animated, and the double-press sets maskPath, maskOpacity, timeRemap, expressions, material, paint, selected, missingEffects, waveform), timeline.keyAt.<anchor|position|scale|rotation|opacity> (Alt+Shift+A/P/S/R/T), timeline.twirlSelected (Ctrl/Cmd+), timeline.collapseAll, flowchart.options {layers?, effects?, solids?, direction?: lr|tb, comp?}, flowchart.graph {comp?}→{nodes, edges}, window.workspace {name}` (all After Effects workspaces).
engine.commands {filter?, enabledOnly?} [{id, label, menu, shortcut, params, enabled, why}]
ui.menu.list {} The menu tree with enablement.
render.frame {comp?, time?, max_side?, path?, base64?} Renders a comp frame through the session, independent of the viewer zoom or resolution. comp is an id or name (default: the active comp), time is in comp seconds (default: the CTI), max_side caps the longest side (0 = full size). Writes a PNG to path (default: a temp file) and returns {comp, time, width, height, path}. With base64: true it returns {…, png: "<base64>"} and writes no file.

Useful query commands for engine.execute (they aren't journaled or undoable): project.summary, comp.info {comp?}, layer.tree {layer, comp?, depth?, time?} (every node has a path), prop.get {layer, path, comp?, time?}, editor.state, command.list {filter?, enabledOnly?, schemas?}, command.describe {command} (params doc and JSON Schema), app.capabilities (version, counts, export formats, parity summary), jobs.list (Progress panel jobs, e.g. the background footage.check after File ▸ Open).

Every engine command is reachable through engine.execute, and every interactive widget of every panel registers an element id; both are enforced by crates/ui-egui/tests/ui_agent_audit.rs.

UI state

Method Params Result
ui.inspect {} Window size, pixels per point, fps, serialized UiState, playback, time, active comp, selection, open dialog, last render ms, keyboardFocus, windowFocused, workspaceStorageError.
ui.elements {prefix?} [{id, label, rect: [x, y, w, h]}]: the widgets drawn in the last frame, in points.
ui.set {tool?, workspace?, theme?, focused?, viewer?, timeline?, menuBar?, home?} viewer: {zoom (null = fit), res: Full|Half|Third|Quarter|Auto, pan: [x,y], grid, rulers, safeMargins, transparencyGrid}; timeline: {pps (null = fit), start, graphEditor, showModes, openLayers: [ids], openGroups: [uids]}.
ui.panel.show / ui.panel.close {panel} Panels: Project, Composition, Timeline, EffectControls, EffectsPresets, Info, Preview, Character, Paragraph, Align and others.
ui.playback {action: play|stop|toggle|status} {playing, time, audio, levelsDb, peaksDb} (audio: an audio preview is running; L/R meter levels and held peaks in dBFS)
ui.resize {width, height} Resizes the window (points).
ui.focus {} Activates the window and takes keyboard focus. Agents normally avoid this, because the app is started without stealing focus.
app.quit {force?} Closes the app. A modified project shows the Save / Don't Save / Cancel prompt first (automation ids dialog.unsaved.save, dialog.unsaved.dontSave, dialog.unsaved.cancel) unless force: true.

keyboardFocus is null when no egui widget has keyboard focus. Otherwise it contains the actual widget id (hexadecimal), its rect: [x, y, w, h] when available, an automationId when the widget registered its response identity, and registryCandidates containing exact rectangle matches. Candidates alone do not establish widget identity. Header, toolbar and dock-tab controls register their response identity. ui.focused continues to identify the focused panel; it is separate from keyboard widget focus. windowFocused reports the native window focus received by egui, which can differ from the retained widget focus.

When a configuration store is available, current dock and floating geometry, focused and maximized panels, and named workspace snapshots persist in workspaces.json. The current layout and named snapshots are independent: changing a divider does not overwrite the named snapshot until window.saveWorkspace runs. Changes save after pointer release, at most once per second, with a final save on normal exit. Version 1 files are limited to 1 MiB and bounded workspace trees; extra viewer IDs must be between 1 and 65,535. Native reads enforce the size limit while reading and report I/O and UTF-8 errors. Unreadable, invalid or newer files leave the default workspace available and are preserved for the entire session, including after later layout changes. Load and save errors appear in the status message and the nullable workspaceStorageError field; a failed write remains eligible for retry. window.saveWorkspace, window.saveWorkspaceAs and window.editWorkspaces (rename, delete) write the file immediately. Through ui.menu.invoke they return an error when it isn't written (a failed write, a file preserved after a failed load, or no configuration store); engine.execute hands them to the UI and returns {frontend} first, so check workspaceStorageError there. The changed workspaces still apply for the session, and a later save retries them.

Input

Points are logical window coordinates. A target is {id} (the element's centre; add fx/fy between 0 and 1 to pick a spot inside it) or {x, y}. modifiers is {shift, alt, command, ctrl}, where command is ⌘ on macOS and Ctrl elsewhere.

Method Params
ui.click {id | x,y, button?: left|right|middle, count?, modifiers?}
ui.move {id | x,y}
ui.drag {from: target, to: target, steps? (12), modifiers?}: press, move, release
ui.scroll {id | x,y, dx, dy}
ui.key {key}, for example Space, F9, Cmd+Z, Cmd+Shift+Z, PageDown, Escape
ui.type {text}: text input into the focused field

Shortcut presets also accept Physical+<shortcut> for a keyboard position, using egui's US-position names (for example Physical+6 and Physical+Minus distinguish AZERTY's - and ) keys). Physical bindings take precedence over logical bindings for that key press. ui.key accepts the same notation to test them. Ordinary shortcuts remain layout-aware. time.workAreaStart and time.workAreaEnd navigate to the first and last included work-area frames and can be assigned through shortcuts.set.

Locating things

Method Params Result
ui.viewer.locate {x, y} (comp pixels) {x, y} window point
ui.viewer.hit {id | x,y} (window point) {x, y} comp pixel
ui.timeline.locate {layer (id), time?} {x, y} window point on the layer bar

Screenshots

Method Params Result
ui.screenshot {path?, panel?, id?} {path, width, height}: a PNG of the window, or cropped to one panel or element.

Element ids are stable, for example tools.Selection, panel.Timeline, panel.tab.EffectControls, panel.tab.Timeline.<comp> (one Timeline tab per open comp; panel.tab.Timeline is the shown one), panel.tab.Viewer(<n>) and viewers.<n> (View ▸ New Viewer's other Composition viewers; viewers.0 is the Composition panel while another viewer is active), header.workspace.Animation, project.item.<id> (a double-click opens the item, Enter renames it), viewer.comp, viewer.handle.<layer>.<i>, timeline.layer.<id>.bar, timeline.layer.<id>.nestedMarker.<i> (a precomp's comp markers), timeline.layer.<id>.twirl, timeline.layer.<id>.row (click selects, Enter or a double-click renames, drag reorders), timeline.prop.<uid>.stopwatch, timeline.key.<uid>.<frame> (and timeline.keyAt.<uid>.<ticks>, exact in layer-time ticks, for keys on subframes), timeline.cti, effectControls.prop.<uid>.value (an angle's revolutions are timeline.prop.<uid>.revolutions, effectControls.prop.<uid>.revolutions and properties.prop.<uid>.revolutions) and effects.item.<name>. The Composition viewer adds viewer.magnification, viewer.resolution, viewer.roi, viewer.grid, viewer.channel, viewer.exposure, viewer.snapshot, viewer.showSnapshot, viewer.fastPreviews (their popup entries are viewer.<menu>Item.<n>), viewer.ruler.top|left|origin, viewer.mask.<uid>.vertex.<i>, viewer.shapePath.<uid>.vertex.<i>, viewer.motionPath.<layer>.<key>[.in|.out], viewer.puppetPin.<uid>[.rotate|.scale], viewer.freeTransform.handle.<i> and viewer.regionOfInterest; the Graph Editor adds timeline.graph.transformBox[.<i>], timeline.graph.snap and timeline.graph.reference. Use ui.elements to see what is on screen.

Viewer state that agents drive headless too: view.snapping, view.snappingOptions {edgesExtended?, collapsedFeatures?, toggle?} (the Tools bar's Snapping options, header.snappingOptions: After Effects' Snap Edges Extended and Snap to Features in Collapsed Compositions and Text Layers), view.channel {channel, colorized?}, view.exposure {stops | delta}, view.resetExposure, view.takeSnapshot, view.showSnapshot, view.fastPreviewMode {mode} (one mode each, for shortcuts: view.fastPreviews.off|adaptive|draft|fastDraft|wireframe, and for the channels view.channel.red|green|blue|alpha|redColorized|greenColorized|blueColorized|rgbStraight), view.res.full|half|third|quarter|auto {comp?} and view.res.custom {factor, comp?} (the comp's Resolution, saved with the project but no undo step; ui.set {viewer: {res}} sets the active comp's, and ui.inspect reports it as ui.viewer.res), view.setRegionOfInterest {rect}, view.addGuide {percent?}, view.addPercentGuide, view.convertGuides {to: percent|pixels, index?}, view.moveGuide (comp pixels, whatever the guide's unit), view.removeGuide, effect.setLabel {layer?, effect, label}; editing: shape.newPath (Pen on shape layers), shape.newShape (the shape tools: a new group in the selected shape layer, else a new shape layer), shape.toolOptions (the Tools bar's Tool Creates Shape / Mask and Fill and Stroke Options), mask.insertVertex, mask.convertVertex, mask.deleteVertices, path.freeTransform (masks and shape paths by uid), keys.setSpatialTangents (motion-path handles) and keys.transform (Graph Editor transform box, timeline Alt-drag scaling).

The MCP server's bridge mode (effectcraft-cli mcp --bridge 9877) is a thin client of this protocol; see agents.md.

Keyframe command times (prop.set, prop.addKey, keys.set, keys.select) default to layer seconds. Set timeBase: "comp" for comp seconds; keys.set.newTime and every keys.select.keys[].time use the same input base. Omitted prop.set or prop.addKey times use the target comp's CTI. Supplied write times snap to its nearest comp frame before conversion to stored layer time, including on stretched or reversed layers. Existing keys.set neighbor constraints and roving adjustment can change the final stored time.

Explicit bases add layerTime and compTime to unsplit prop.set and prop.addKey replies, alongside value or count. Separated Position prop.set keeps an array of per-axis replies. Explicit keys.select replies contain count and keys[], with timing on each key in the resulting selection. Key metadata's time remains layer seconds regardless of the input base. Without timeBase, these three commands retain their existing JSON value/count replies, including outside the layer's [in, out) range. Timing metadata carries outside-range warnings; keys.set and prop.get.keys[] report both bases and warnings without an explicit base too. prop.get.time is comp seconds. For these commands, path or prop identifies a property, not a group. Set a layer's offset with layer.timing {layer, start} (comp seconds).

Panel moves, tab ordering, edge splits and floating geometry use shared docking transactions. window.floatPanel requires exactly four finite numbers when rect is supplied. Invalid geometry or missing source/anchor panels leaves the layout unchanged. The persisted workspace shape is unchanged: accordion open states and explicit heights, floating groups, and dynamic viewer/ScriptUI identities remain app data. Pointer drops use the same shared placement rules as commands, with floating groups resolved in their actual window stacking order.

window.dockAction accepts {action} using the shared docking action's JSON form. It covers tab moves and ordering, split and floating geometry, activation, and closing. For example, {"action":{"ResizeSplit":{"path":[],"size":{"FixedFirst":360}}}} resizes the root split; path entries use false for the first child and true for the second. {"action":{"MoveFloating":{"panel":"Info","rect":[45,65,420,320]}}} repositions and resizes the floating group containing Info. Invalid actions leave the layout unchanged. Pointer gestures dispatch through this same catalog command. EffectCraft also supports SetStackOpen and ResizeStack. Dynamic panel identities use {"Viewer":42} or {"ScriptPanel":77} in the action. Closing runs the existing viewer and ScriptUI lifecycle callbacks; opening uses the existing panel creation commands.