Be Your Evolution AI News Daily

1. Sync Labs Agent Skills

Sync Labs has packaged ten video operations as portable agent skills: lip sync, talking photos, dubbing, dialogue repair, campaign localization, batch processing, troubleshooting, and API integration. The repository includes client manifests, translated guides, validation scripts, tests, and mocked MCP evaluations rather than a few prompt files. It is especially relevant when an agent must plan and resume many video jobs. The skills also quote cost before paid work and require approval before proceeding.

Operational value: Give Codex or another skills-compatible agent a documented workflow for dubbing and lip-sync jobs, including batching, resumability, cost confirmation, and failure recovery.

Quick video: Dub one short talking-head clip into Spanish, show the quoted cost and approval pause, then compare the original mouth timing with the generated result.

Setup / license reality: The skill repository is MIT-licensed, but it connects to Sync’s proprietary hosted MCP and video APIs, requires an account, and may incur usage charges. Browser sign-in avoids copying a key for interactive use; scripts and CI need an API key. Media, voices, translations, likeness rights, and generated outputs remain separately governed. Treat uploaded footage as data sent to a third-party service.

2. Architecture Diagrams

Architecture Diagrams is an agent skill that converts a codebase, infrastructure description, or plain-English brief into an editable draw.io file plus SVG or PNG. Its method is the useful part: resolve real vendor icons, plan a grid, build from a YAML or JSON specification, lint overlaps and crossings, render a preview, and visually review it. The bundled catalogs cover major clouds, Kubernetes, Microsoft Fabric, and thousands of brand marks, while missing icons fail visibly.

Operational value: Turn a repository or deployment plan into an editable system diagram with consistent layout rules, real service icons, automated geometry checks, and retained source.

Quick video: Ask the skill to read a small Terraform project, watch it resolve cloud icons, then open the final draw.io file and move one component manually.

Setup / license reality: MIT-licensed skill and code. Python 3.10+ is required; YAML specs need PyYAML, and Chrome, Edge, or Chromium is used for PNG previews. Internet access is needed the first time icons are fetched. Vendor icons retain their owners’ terms and should describe services without implying endorsement. Native AWS and Kubernetes stencils require draw.io Desktop or internet-backed rendering.

3. AlwaysCut

AlwaysCut is an ambitious local-first video editor built around an editable multitrack timeline rather than a one-shot generated clip. Its repository describes transcript-based cuts, captions, effects, motion graphics, keyframes, subject-aware reframing, Remotion export, MCP access, and agent skills for Codex and Claude Code. Projects and media stay local by default, while configured model or generation services may still receive requests. The feature breadth is compelling, but the public repository was created today and needs real-world maturity.

Operational value: Let an agent assemble and revise a real timeline while preserving project data, media tracks, captions, effects, exports, and undoable editor commands.

Quick video: Import a talking-head clip, ask the agent for a rough cut and captions, then drag one edit manually and export the retained timeline.

Setup / license reality: AGPL-3.0 code, not permissively licensed. Desktop installers and source-build instructions are documented, but this is a brand-new project with a very large surface area. Local-first does not mean every AI feature is offline: provider calls, media search, and generation depend on what you configure. Keep MCP loopback-only, protect any public endpoint with a bearer token, and test only on copied media.

4. ComfyUI-LyricSRT

ComfyUI-LyricSRT turns audio and known lyrics into timed SRT subtitles by combining Qwen3-ASR word timing with global text alignment. That means the subtitle text stays exactly as supplied even when recognition mishears a sung word, and repeated choruses remain ordered. A second node provides audio playback, line-level start and end adjustments, text edits, speed control, shifting, deletion, and local Save As. It is a narrowly focused tool with obvious value for music-video finishing.

Operational value: Create a local, editable subtitle track from a finished song and known lyric sheet, then correct timing inside the same ComfyUI workflow before export.

Quick video: Feed one chorus into the node, watch exact lyrics align to the vocal, adjust two line boundaries by ear, and save the corrected SRT.

Setup / license reality: MIT-licensed node code. The documented setup needs ComfyUI, Python packages, Qwen3-ASR-1.7B, and Qwen3-ForcedAligner-0.6B; both model cards identify their weights as Apache-2.0. The author tested ComfyUI 0.44 and frontend 1.59 on Windows. Models consume local storage and GPU memory, and the editor’s Save As dialog appears on the machine running ComfyUI. Always review automated lyric timing by ear.

5. tasq

tasq is a local-first task tracker designed for several coding agents working in one repository. A single Rust binary and SQLite file model nested tasks, expiring work leases, file or directory reservations, dependencies, checkpoints, deadlines, and burn-up. MCP runs over standard input and output, while a human CLI exposes trees, boards, diagnostics, and progress. The design directly addresses duplicate work and edit collisions, but the repository is days old and its detailed primary documentation is currently Russian-first.

Operational value: Give concurrent agents durable task ownership, expiring claims, dependency visibility, handoff checkpoints, and conflict-aware reservations without adding a cloud account or server.

Quick video: Let two agents claim neighboring tasks, trigger a reservation conflict on the same file, then release one lease and show the clean handoff checkpoint.

Setup / license reality: MIT-licensed Rust code with SQLite and an MCP stdio server. Version one intentionally omits a web UI, network synchronization, accounts, roles, and cloud telemetry; HTTP is loopback-only behind an explicit flag. Those limits are sensible, but the project is new, has no broad adoption evidence, and should be tested on a disposable repository before trusting its coordination guarantees.

MIT-licensed LangGraph remains a strong foundation for stateful, long-running agents that need explicit graphs, durable execution, human approval points, memory, streaming, and controlled recovery from interruptions.

MIT-licensed MLX remains Apple’s useful array framework for local machine learning on Apple silicon, with NumPy-like APIs, automatic differentiation, lazy evaluation, and shared CPU-GPU memory.

Apache-2.0 LLaMA-Factory remains a broad fine-tuning workbench for many language and multimodal models, spanning supervised tuning, preference methods, quantization, distributed training, inference, and a web interface.

Mastra remains a capable TypeScript framework for agents, graph workflows, memory, MCP, evaluations, and observability; its core is Apache-2.0, while directories named ee use a separate enterprise license.

Apache-2.0 Keras remains an approachable high-level deep-learning API whose current multi-backend design can run models with JAX, TensorFlow, PyTorch, or OpenVINO inference instead of locking code to one stack.