Skip to content

RESEARCH.md

Latest commit

 

History

History
89 lines (79 loc) · 10.2 KB

RESEARCH.md

File metadata and controls

89 lines (79 loc) · 10.2 KB
Raw

Research — OpenCut

Executive Summary

Verified: OpenCut v1.33.1 is a local-first Adobe Premiere Pro automation stack: Python/Flask on localhost, CEP plus UXP panels, generated route/readiness/model/MCP manifests, FFmpeg/model-backed media workflows, plugin trust controls, and local installers. Its strongest current shape is not "another video editor"; it is trustable local automation that keeps capability, privacy, provenance, and recovery state explicit before an editor runs a workflow. Highest-value opportunities: repair the current i18n and job-resume verification failures; stop leaking remote-render-node secrets; reconcile real social upload routes with stub publish routes and README claims; keep MCP/model generated docs in lockstep with manifests; retire WebSocket/FastAPI "coming soon" surfaces or wire them to the real bridge; and keep UXP/live Premiere migration evidence in the blocked queue until a host run exists.

Product Map

  • Core workflows: Premiere panel automation for cuts, captions, audio cleanup, VFX, Magic Clips/shorts, social/export presets, timeline write-back, plugin actions, model readiness, and REST/MCP automation.
  • User personas: Premiere editors who want local repetitive-work automation; creators avoiding cloud usage caps; post/ops users who need batch media tooling and recoverable jobs; technical users scripting the REST/MCP/plugin surfaces.
  • Platforms and distribution: Python 3.11+ server on Windows/macOS/Linux, CEP panel, UXP migration path, WPF Windows installer, Docker/local server, MIT opencut-ppro package metadata.
  • Key integrations and data flows: CEP/UXP -> Flask routes -> async jobs/history -> FFmpeg/model backends -> generated route/readiness/model-card/MCP artifacts; plugins load from ~/.opencut/plugins; optional social, cloud, telemetry, and C2PA paths are explicit opt-ins.

Competitive Landscape

  • Adobe Premiere Pro / UXP / Content Credentials: strong native AI search, transcripts, captions, UXP distribution, and provenance expectations. Learn: UXP and content authenticity are trust basics. Avoid: claiming migration completion without live Premiere UDT evidence.
  • FireCut, AutoPod, TimeBolt, AutoCut: paid Premiere automation wins by making common edits obvious and reliable. Learn: command surfaces must lead to ready actions, not stubs. Avoid: mixing export-prep stubs with real upload flows under the same label.
  • Descript, OpusClip, Submagic, Gling: cloud creator tools lead on transcript editing, caption polish, short-form handoff, templates, and publishing. Learn: social handoff and caption readiness must be honest and polished. Avoid: cloud-default collaboration that weakens OpenCut's local-first privacy value.
  • LosslessCut and auto-editor: OSS tools succeed through focused local FFmpeg-backed behavior and deterministic CLI flows. Learn: local reliability beats broad unverified capability lists. Avoid: large catalogs with stale counts or dead roadmap anchors.
  • Kdenlive and Shotcut: local NLEs show mature export/status/media handling. Learn: errors, export state, and media interchange need careful UX. Avoid: trying to become a full NLE inside Premiere.
  • OpenTimelineIO, Remotion, Editly: adjacent tools prove the value of schema/versioned interchange and reproducible generated video. Learn: generated manifests and stable contracts are worth preserving. Avoid: hand-maintained docs where generated data already exists.
  • OpenCut-app/OpenCut and pyvideotrans: public issue signal emphasizes API/headless workflows, auto-import, multilingual handling, and install reliability. Learn: local multilingual automation is a real demand. Avoid: duplicating a browser editor product shape.

Security, Privacy, and Reliability

  • Verified risk: py -3.12 scripts/i18n_lint.py --check fails with 51 consumed-but-missing CEP keys, including palette.* and settings.plugin_*; focused tests/test_i18n_drift.py also fails, so recent plugin/palette UI strings are not release-ready.
  • Verified risk: tests/test_job_resume.py::test_checkpointable_routes_are_marked_resumable fails because the test searches for an exact decorator string, while opencut/routes/video_specialty.py correctly has @async_job("shorts_pipeline", resumable=True, rate_limit_key="ai_gpu").
  • Verified risk: opencut/core/remote_process.py persists remote-node api_key values in ~/.opencut/remote_nodes.json, and opencut/routes/remote_realtime_routes.py returns node.to_dict() from register/list endpoints, exposing raw node secrets to any authorized local caller.
  • Verified risk: remote render-node registration accepts arbitrary URL strings and immediately calls <url>/health; OWASP SSRF guidance recommends strict scheme/host validation, no embedded credentials, explicit private-network policy, and tests for resolver edge cases.
  • Verified risk: /social/upload delegates to real OAuth upload logic in opencut/core/social_post.py, but /publish/upload uses opencut/core/multi_publish.py::publish_to_platform(), which is documented as a stub returning pending_upload. README and command palette language can overstate direct publishing depending on which surface a user enters.
  • Verified drift: docs/MCP_SERVER.md and opencut/routes/mcp_bridge_routes.py still cite 39 curated tools and 1,325 extended tools while the changelog says 86 curated tools and opencut/_generated/mcp_extended_tools.json reports 1,450 tools.
  • Verified drift: opencut/model_cards.py, docs/MODELS.md, and generated feature/model manifests still contain install hints such as stub — roadmap H2.1 even though the live root roadmap no longer has those anchors.
  • Verified risk: opencut/core/fastapi_app.py exposes a /ws/jobs WebSocket stub that replies "WebSocket job updates coming soon" while opencut/core/ws_bridge.py contains a real WebSocket bridge used by the app.
  • Verified i18n quality gap: py -3.12 scripts/lint_locales.py returns success but warns on Spanish UXP missing diacritics, so locale quality can regress without breaking release checks.
  • Missing guardrails: release smoke should cover the i18n drift and generated-doc count failures that currently escape narrower manifest checks.

Architecture Assessment

  • opencut/core/remote_process.py needs a public/private serialization boundary: persisted secrets may remain local, but route responses and logs should use redacted dictionaries and URL validation helpers.
  • The publish stack has two competing meanings of "upload": real OAuth upload in social_post.py and dry-run/export-prep in multi_publish.py. Unify route naming, responses, and panel/command copy so users know when a platform API call actually happened.
  • Generated data already exists for MCP tools, model cards, feature readiness, and route readiness; docs should read from those generators or have failing tests for stale counts and dead anchors.
  • tests/test_job_resume.py should parse decorator metadata or route registration state instead of brittle exact source strings; the current source contains the correct resumable marker.
  • FastAPI migration code should either bridge to the real job WebSocket mechanism or keep the stub unavailable from supported docs/tests until it is truthful.
  • Category coverage: security maps to remote-node SSRF/secret redaction; accessibility and i18n map to CEP missing keys and UXP locale warnings; observability maps to job WebSockets and release smoke; testing maps to drift gates; docs and distribution map to generated MCP/model counts and blocked package-manager releases; plugin ecosystem is shipped enough to need locale completion; mobile and cloud multi-user are product misfits; migration remains evidence-gated in Roadmap_Blocked.md.

Rejected Ideas

  • Cloud-first collaboration from Descript/OpusClip/Submagic: rejected because OpenCut's differentiator is local-first Premiere automation; keep exports/review handoffs optional.
  • Mobile app: rejected because the product is a desktop Premiere/server workflow.
  • Full NLE replacement: rejected because Premiere is the host; Kdenlive/Shotcut are references for reliability, not target product shape.
  • More speculative AI/model stubs: rejected until existing stub/model-card readiness docs stop pointing at dead roadmap anchors.
  • GitHub Actions as the release/test answer: rejected because repo policy and current .github state are local-build only.
  • Directly implementing PyPI/Homebrew/winget in this pass: rejected because those are already credential/signing-gated in Roadmap_Blocked.md.
  • Moving UXP WebView cutover back to active work: rejected until live Premiere UDT evidence exists.

Sources

Repository and OSS:

Commercial and platform:

Standards, APIs, dependencies, and security:

Open Questions

  • Which human-controlled signer/key workflow should be canonical for release-grade C2PA credentials?
  • When can a real Premiere 26.x UDT capture be produced for window.OpenCutUXPUdtHarness.run({ includeMutating: true })?