palmier-io/palmier-pro

Palmier Pro is a single macOS Swift application organized as a set of feature modules under one app target. It is a native video editor, not a web app, and it runs locally on Apple Silicon with a UI built from SwiftUI/AppKit-style views, view models, and window controllers.

The top-level app shell lives in App/ and is responsible for lifecycle, global app state, notifications, menu wiring, and update handling. main.swift boots the app, AppDelegate and AppState coordinate startup and shared state, and Updater plus Changelog manage in-app release/update behavior.

The core user workflow is project-centric. Project/ handles home/welcome screens, project cards, registry, sample projects, and project settings mismatch handling. Models/VideoProject.swift and related model types define the persisted project structure that the editor, exporter, agent, and MCP server all operate on.

The main editor surface is split into three major interaction zones: the timeline, preview, and inspector/media side panels. Editor/, Timeline/, Inspector/, MediaPanel/, Preview/, and Toolbar/ together form the editing workspace, with EditorViewModel.swift acting as the central state coordinator for tracks, clips, keyframes, clipboard, ripple edits, linking, relinking, project settings, and preview tabs.

The timeline is where most editing state changes happen. Components such as TimelineView, ClipRenderer, SnapEngine, OverwriteEngine, and RippleEngine handle clip placement, snapping, drag state, range selection, and destructive vs. ripple-style edits. The preview subsystem renders the current timeline state through CompositionBuilder, TimelineRenderer, VideoEngine, and overlay views for crop/transform/text interactions.

Media is modeled separately from timeline edits. Models/MediaAsset, MediaFolder, MediaManifest, and MediaResolver describe imported content and how it is found on disk; MediaPanel/ provides browsing, search, drag-and-drop, and folder tiles; StoragePane and related settings govern local storage behavior. This boundary keeps asset management distinct from timeline composition.

AI generation is a first-class subsystem. Generation/ contains model catalogs, cost estimation, backend selection, and submission types for image, audio, music, and video generation. GenerationService and the submission types bridge the editor UI to a backend, while Generation/Edit covers AI-assisted edit actions such as trim extraction and submitter logic. The README makes clear that the generative processing itself is the closed-source part of the product.

The in-app agent is another major subsystem. Agent/ provides the chat UI, session storage, mention handling, prompt/tool orchestration, and client implementations. It supports two external agent backends through AnthropicClient and PalmierClient, and the agent panel is wired into editor/project state so it can manipulate the same timeline the user sees.

Palmier Pro exposes its local editor to external agents through an MCP server boundary. Agent/MCP/MCPHTTPServer.swift and MCPService.swift serve an HTTP MCP endpoint on 127.0.0.1:19789/mcp, and the app bundles a Claude Desktop extension manifest under Resources/MCPB plus a separate mcpb/server package. This means external tools like Claude Code, Codex, Cursor, or Claude Desktop can call into the app while it is running.

The agent’s capabilities are implemented as explicit tools rather than free-form access. Agent/Tools/ defines the tool schema, instructions, results, and concrete tool executors for searches, captions, folders, timeline inspection, text extraction, import/generate operations, and short ID helpers. These tools operate on editor/project state and are the bridge between natural-language requests and deterministic timeline mutations.

Search and transcription are separate local intelligence layers. Search/ builds an embedding-based visual search index using a model downloader, frame sampler, visual embedder, and query engine; Transcription/ provides transcript caching and transcript search. These systems feed both the media panel and the agent/tooling so clips can be found semantically, not just by filename.

Export is a distinct output boundary. Export/ includes a project exporter and XML exporter, and tests indicate round-tripping plus composition generation are important guarantees. This suggests the app can serialize the project for interchange and also render or package it for downstream video workflows.

Account and backend integration are isolated from the editor core. Account/ manages login, credit summaries, top-off, backend configuration, and identity/account UI; GenerationBackend and the agent clients are the main consumers of this configuration. In practice, the free editor path is local-first, while AI generation requires authenticated backend access.

Cross-cutting support lives in utility and infrastructure modules. Utilities/ provides disk caching, keychain storage, logging, async throttling, formatting, and bundled font support; Telemetry/ tracks usage; Help/ and Settings/ expose onboarding, shortcuts, privacy, model, storage, notification, and agent configuration. scripts/, mcpb/, and .github/workflows/ci.yml show the project is built and packaged as a distributable macOS app with CI, release automation, and MCP desktop-extension packaging.

Overall, the architecture is a local native editor with three major external boundaries: a closed generative AI backend, a local MCP server for agent integration, and the user’s local filesystem/media library. Most feature modules feed one shared project/timeline state, which is the central data model that everything else reads from or mutates.