Roadmap

If you want to help, the Now section is where help changes the most.

Where it stands today#

Working and used daily: compositions and nesting, 2D/3D layers with cameras and lights, keyframes and the graph editor, 183 effects, masks and mattes, shape layers, per-glyph text animators, expressions, bone and puppet rigging, particles, SVG and Lottie import, and export to mp4/mov/webm/GIF/PNG/Lottie through a local ffmpeg.

Added 2026-08-17: a named ease-curve library, a disk tier under the frame cache, onion skinning, text and colour Essential Properties, cloners with effectors and layer-driven fields, and 2D rigid-body physics.

Added 2026-08-18: the frame cache is keyed on scene content and survives undo and restart; output-module templates; cloner cascade, push and path modes; and physics rotation (opt-in per body — real OBBs, contact torque, rolling).

Added 2026-08-30: a headless CLI (premation render) over the same deterministic pipeline the editor exports with; data-driven batch rendering, one file per row of a CSV; captions — .srt/.vtt in and out as ordinary text layers, plus generation from the composition's own audio; auto-reframe to another aspect ratio, following the subject and jumping at cuts; the pick-whip, for parenting and for expressions; a download-on-demand installer for the Object Matte model; and idle caching of the whole work area rather than five seconds ahead.

Added 2026-09-02, in three passes over an audit of the editor. Most of what it found was not missing engine work but finished engine halves with no control, so a lot of this is reach rather than build.

The footage path became an editing loop: a Source Monitor (in/out in source seconds, JKL shuttle, Insert / Overwrite / Add to end / New comp from range), timeline edit tools as visible modes — Selection, Razor, Slip, Slide and Roll on Shift+S/C/Y/U/R, with roll being a real two-sided trim that did not exist before — clip-edge snapping and Fit Composition / Fit Work Area, per-cut transitions held as records you can select, lengthen and remove, Assemble from Footage and New Composition from Selected Clips, a Scopes panel (waveform, parade, vectorscope, histogram), a Transcript panel with text-based editing, silence removal and ducking, chapters from labelled markers on MP4/MOV export, and a render queue that pauses and resumes (within a session) with Discard as the separate destructive verb.

Motion got one graph editor — the Motion panel's private copy is gone, and the survivor carries Animated / Selected modes, a frozen reference curve and a saved-curve ease library — plus parametric stagger with a Re-apply that replaces rather than layers, modifier stacks (ordered offset / multiply / clamp / wiggle / smooth / spring / loop / delay / audio / oscillate rows compiled to one expression), audio-reactive drivers, bake dynamics to keyframes, The Smoother and The Wiggler as real dialogs, Keyframe Velocity and an Interpolation submenu, onion-skin settings, and caret autocomplete in the expression editor.

3D got image skies from any image or EXR asset and, with them, reflections — a prefiltered specular atlas and split-sum IBL in both shader dialects; the full glTF PBR map set (normal, metallic-roughness, occlusion, emissive) and external .gltf files with their sidecars, behind a File ▸ Import 3D Model entry; real curved primitives (sphere, cylinder, cone, torus, capsule, box) as meshes with editable segment counts; a material editor and persisted library; Composition Settings ▸ World; light presets and a Kelvin row; bevel styles; morph-target sliders; 3D IK on parented layers; and the gizmo and DOF focus plane working in every 2-up / 4-up pane.

And around the edges: a knife tool and pathfinder for shape paths, an on-canvas gradient editor, smart guides with distance badges and equal-spacing detection, project swatches, an interactive onboarding tour, one home for the preview controls (with Cache Work Area Now, Purge RAM and Purge Disk), and Window ▸ Workspace.

The engine is one GPU render graph (WebGPU, falling back to WebGL2) shared by the viewport and the exporter, covered by golden-image render tests.

Pre-1.0. Expect rough edges in the UI and occasional breaking changes to the .motion bundle format before 1.0.

Now#

A local project browser — SHIPPED 2026-08-20#

The start screen is now a real project browser: a card grid backed by the local index (comp facts, save revision, thumbnail), joined with the MRU for anything the index has not seen. What made it possible was less UI than plumbing — the index had a full API, an IPC bridge and tests but ZERO writers, so it had been empty since it shipped. Saves and opens now write it (indexWriter.ts), a LocalThumbnailWorker captures project thumbnails to a content-addressed cache dir (thumb:* IPC), and the thumbHash column the SQLite schema always had finally gets values. In a browser tab it degrades to the MRU list, as before.

Finish verifying local-first on-device#

ASSETS are done. A bundle now COLLECTS session-local footage on save — the bytes content-addressed into blobs/, recorded in assets/registry.json, and the document repointed at them — and RESTORES that registry on open, so a project moved to another machine renders the footage it was authored with and its Assets panel comes back populated.

Verified end to end by rendering a hand-built bundle whose document carried nothing but a dead blob: URL and an asset id, and by the control: with the registry removed, the export refuses with "media offline" rather than shipping a frame with a hole in it.

The rest of the disk-backed paths have still not been exercised end-to-end on a real device. Specifically: better-sqlite3 needs an electron-rebuild against the Electron ABI, and the binary blob IPC needs a real save/load cycle. Concrete, verifiable, and a good way to learn the storage layer.

Next#

• A real video decoder — subsystem AND renderer swap SHIPPED 2026-08-19. src/core/video/ is the WebCodecs path: mp4box demux (pure JS, so the demux and the GOP/B-frame index are jest-pinned against real ffmpeg fixtures), a frame index answering "which samples decode frame N", and ExactVideoSource (feed key→target, flush, cache the GOP). Consumers: the footage preview's Frame-by-frame mode, and the RENDER PATH — exactVideoFrames.ts serves exact frames to viewport and export first, with the element-seek path (videoFrameCache) as the permanent fallback tier for WebM/odd containers, oversized files and WebCodecs-less runtimes. Export exactness rides the existing takeMediaWaits convergence loop. Real-machine visual pass: DONE the same day, in a real Chromium tab — frame-by-frame stepped an ffmpeg fixture with pixel readback, and the render path's cache reported all 24 frames decoded for the composited clip. The tracking column has since landed on top of it — see the next bullet.
• After Effects parity. The tracking column SHIPPED 2026-08-20, all four modes: Track Motion on video layers (src/core/tracking/ — NCC matching with Lucas-Kanade sub-pixel refinement, anchor-template drift correction, occlusion coasting) runs on the exact decoder, walks any number of points through one decode pass, and applies as Follow (position keyframes on any layer), Stabilize (inverse motion on the footage), Corner pin (keyframes a Corner Pin effect on a target — screen replacement), or Track mask (every mask vertex tracked, written as mask keyframes — rotoscoping's first step), and Follow + rotation & scale (2-point solve: anchor drives position, the anchor→reference vector drives rotation and scale, angle-unwrapped past ±180°). Still open on the column: a roto brush / edge-aware masks — though the ROTO BRUSH's neural half is now reachable, because the piece that was missing was never code. The segmenter, the ONNX wrapper and the runtime have all been in the tree; what nobody had was a way to GET a model, since a build-time environment variable is not something a person does. Settings ▸ Object Matte downloads one on an explicit press and caches it for every later launch. The PICK-WHIP shipped too, for parenting and for expressions — the most-felt gesture gap. Beyond tracking: more of the effect set, richer expression bindings. Open an issue for the specific gap you hit — that is far more useful than a general "more parity" wish.
• A video editing loop. SHIPPED 2026-09-02 — source monitor, edit-tool modes (razor / slip / slide / roll), clip-edge snapping, per-cut transitions, Assemble from Footage, scopes, transcript-based editing, silence removal and ducking, and render-queue pause/resume. See "Where it stands today".
• Timeline and graph-editor polish. The two graph editors became one on 2026-09-02, with Animated / Selected visibility modes, a frozen reference curve and an ease library — which was most of what "not yet pleasant for dense compositions" meant. What is still open is density rather than capability: the panel is not yet fast or legible with hundreds of tracks on screen.
• Performance on large projects. The engine handles high layer counts, but the UI has hot spots under heavy scenes. Profile before optimising, and bring the profile to the PR.
• Broader render-test coverage. Subsystems with only unit tests can regress visually without anything going red. The harness is in packages/render-tests/.
• Colour management. Linear working-space and linear RT storage shipped; display-referred upload tagging (displayReferred, rgba8unorm-srgb backend support, HARDWARE_SRGB_UPLOADS kill switch default off) shipped; project working-space UI (Composition Settings → Color), ACES display ODT, and 32-bpc intermediate depth shipped (2026-08-14). HDR output still open.

Later#

• Audio. Waveform display, audio layers, offline mixdown, dB automation, per-clip effects and the audio expression all exist (src/core/audio/). What does not: a mixing console — buses, sends, and live metering.
• Plugin ecosystem. The sandbox, permission model and signing all ship. What is missing is discovery outside the hosted registry, and more host API surface for plugin authors.
• Collaboration. Real-time multiplayer would need a substantial re-architecture. Not planned, not refused.
• Advanced 3D. Almost all of what this entry listed as open closed on 2026-09-02. Imported glTF/GLB models (embedded and external-file, with sidecars), CPU skinning, morph targets, baked animation clips, 3D IK, an SH-probe Environment Light fed by any image or EXR asset, split-sum reflections off that same sky, the full glTF PBR map set, and real curved primitives all ship — see docs/3d-layer-model.md for how they map onto ordinary layers. Shadow maps shipped too — opt-in per light, PCF-filtered, byte-identical when off (packages/renderer/src/rendergraph/passes/shadowMap.ts); one mapped light per run, point lights along their aim. What remains is SSAO (blocked: every 3D run draws into a multisampled target, so neither backend can sample its depth — it needs a linear-depth prepass bound before the run draws) and height displacement (not started). The target remains AE's Classic 3D, plus extrusion/bevels and this imported-model foothold — full parametric Advanced 3D is not planned.
• Local AI conversation persistence. The assistant ships in the local edition (BYOK via OS keystore + aiProxy, including ai:image). Thread history still only persists when a backend is present; in-session history works. Wiring aiBundleIO into the chat UI would close the gap.

Not planned#

• A web version. The app depends on the filesystem, native menus and a local ffmpeg. The renderer runs in a browser tab for development, but that is a development convenience, not a product.
• Re-implementing the hosted backend in this repo. Accounts, billing and the encrypted sync vault are a separate closed service. The local edition is designed to need none of them for editing — BYOK AI talks to the user's provider directly from the desktop shell.

How this list changes#

Roadmaps drift. If something here contradicts the code, the code wins and a PR fixing this file is welcome. Priorities move based on what people actually contribute — an area with a motivated contributor beats an area with a plan.