Beav: a local-first AI workbench for Xiaohongshu content operations

The problem Beav is aimed at

Content operations for a Xiaohongshu account is a loop of repetitive chores. You scan feeds for material, save posts and comments, hunt for an angle, draft a caption, produce images in a consistent style, and if the post is a talking-head video, cut the redundant takes and burn in subtitles. Each step tends to live in a different tool. The README frames the gap bluntly: general-purpose AI tools such as Codex or WorkBuddy target programmers or office work, and do not double as a long-term asset library plus a continuing production bench. Beav's answer is to keep the assets and the production surface in the same place. The stated audience is the self-media operator: someone running a brand or creator account, not a developer looking for a scraping library.

What actually runs on your machine

Beav is a TypeScript desktop application distributed for macOS, Windows and Linux, with installers published through GitHub Releases and a separate download page hosted at redbox.ziz.hk for users on mainland Chinese networks. The repository also ships a Chrome and Edge extension for web capture. Data is organised around workspaces: the quick-start instructs you to create one workspace per account or brand, and the trust section describes materials, drafts and projects as being organised around a local workspace. That is the architectural claim worth taking seriously. Collection, drafts and media projects are not primarily server-side objects; they sit in a local store, and model calls are the part that leaves the machine. The README notes that assets, extensions, update artefacts and signatures are all published via GitHub Releases, which is a verifiable distribution path rather than a claim about the code itself.

Collection, research and the daily briefing

The capture path has two entry points. The browser extension saves web pages, social posts and comments into the knowledge base, and the screenshots show Xiaohongshu notes and their comment threads being collected. Separately, an automated research feature searches public content across multiple social platforms, filters candidates, then reads the high-value ones in more depth, with results deposited into the knowledge base for later retrieval. Creator subscriptions refresh supported profiles daily and sync new posts into the same store. The README is honest about the limits here: which platforms work and what gets scraped depends on the current data sources and configuration. Treat the subscription feature as a configured pipeline with a variable success rate, not a guaranteed mirror of any platform. Comment-thread capture is the more distinctive piece, since comment text is where audience objections and phrasing live, and it is rarely captured by generic bookmarking tools.

The media workshop and talking-head editing

The 2.8 release notes describe a video workflow built around word-level transcription, semantics and pause analysis. Beav proposes cuts for repeated takes, filler words and awkward joins, and the base edit can be applied to the timeline. From there you inspect the edit log, edit subtitles directly, and add picture-in-picture, animated captions and camera motion. Optional vocal noise reduction can be toggled so you can compare before and after. The video project opens in its own window with a multi-track timeline, canvas transforms, keyframes, transitions, colour work, masks and audio adjustment, plus an AI assistant panel whose edits enter the undo history. Export lets you choose a target directory, shows render progress and failure reasons, and can emit SRT or WebVTT subtitles on their own. One detail in the README deserves emphasis: picture-in-picture placeholders represent footage you have not supplied yet, and animation plans must be confirmed before they become finished frames. The documentation explicitly says not to treat a plan as a completed shot. That is an unusually candid warning about AI-generated timelines.

Images, templates and the small utilities

Image work happens in a dedicated workspace that manages reference images, aspect ratio, generation progress and results. Templates carry editable text variables and designated reference images, and you can upload an image for the AI to decompose into a reusable template. Creation and submission are separate operations, and online generation depends on account permissions and model configuration. Alongside this sit four utilities that open from the home page: a subtitle extractor, an image metadata cleaner that inspects and strips removable PNG, JPEG and WebP file metadata while writing a separate copy, a still-to-Live-Photo converter with push and pan motion, and the account planning tool. Two constraints are stated plainly. The metadata cleaner does not remove logos or pixel watermarks. The Live Photo tool currently supports macOS and Windows only, Linux is unsupported, and the README says Windows hardware testing and phone import are still pending acceptance, with paired files required to survive transfer.

Running it, and connecting an external agent

The quick-start is short: download the installer from the download page, create a workspace per account or brand, then open Settings, AI and either use the official AI service or enter your own Endpoint, API Key and model. If you want web capture, fetch the Chrome or Edge extension from the latest Release. From there you start from a topic or the daily briefing on the home page, or open the media workshop for existing video. There is also an agent plugin path. Beav Creator lets Codex Desktop or WorkBuddy connect to the local Beav instance over MCP and hand creation tasks to Beav's internal agent, while you review, approve and edit results in the Beav UI. CLI users start the app with beav open, then send a prompt to the host agent that points it at beav.ziz.hk/agent or beav.ziz.hk/workbuddy to install the plugin. The README does not document the MCP tool surface itself, so what the host agent can actually do once connected is not specified in the material available.

Where Beav is the wrong tool

The licence is the first real problem. The repository metadata reports NOASSERTION, while the README badge reads MIT-NC. Those are different statements, and MIT-NC is not an OSI-approved licence; a non-commercial restriction is a restriction, whatever the badge says. If you intend to embed any part of Beav in a commercial product, or to redistribute it, resolve this before writing code against it. This is a factual gap in the repository, not legal advice. The second limitation is platform scope. The whole design assumes Chinese social platforms, especially Xiaohongshu. If your channels are elsewhere, the collection and comment-insight features lose most of their value and you are left with a general writing and video tool. Third, the auto-research and subscription features depend on external data sources and configuration, so their reliability is not something the README commits to. Fourth, the Live Photo utility is incomplete on Windows by the project's own admission. Finally, the README states that installer and public source may not be in sync, which means reading the repository is not a reliable way to know what the binary you install will do.

How it differs from assembling your own stack

The obvious alternative is composition: a scraper such as one of the xhs-downloader projects for capture, a note-taking app or Obsidian vault for storage, a chat client for drafting, and a conventional editor such as DaVinci Resolve or CapCut for the video pass. The README even lists Obsidian among the supported local sources, so the two are not mutually exclusive. The difference is where state lives. A composed stack gives you interchangeable parts and no vendor assumptions, but every handoff between capture, notes, drafting and editing is manual. Beav's bet is that the handoffs are the expensive part, and it pays for that by binding you to its workspace format, its model configuration and its update cadence. If you already have a working Obsidian and Resolve pipeline you are happy with, Beav mostly adds a collection layer and an AI drafting layer, and you should evaluate it on those two alone.

Maintenance, release cadence and what to verify

The release history shows a fast cadence: v2.7.19 on 3 September 2026, v2.7.20 on 4 September, and v2.8.0 on 10 September, the same day as the last push to main. Frequent minor releases with a 2.8 feature drop in between suggest active development, and the README points to a CHANGELOG and ROADMAP for the detail. That cadence cuts both ways. You get fixes quickly, and you also get interface churn; the README warns that entry points may differ between the installer and the public source, so a tutorial written against 2.7 may not match 2.8. There is a Bilibili video tutorial linked, and the app can answer questions about its own usage guide from within a conversation, which is a sensible mitigation for a product that changes this often. The licence question remains the item to settle first, followed by confirming that the specific 2.8 capabilities you need are present in your downloaded build.

Editorial conclusion

Adopt Beav if you run one or more creator accounts and want collection, drafts, images and talking-head edits inside a single local workspace rather than scattered across browser tabs and chat windows. Do not adopt it if you need a permissively licensed component to embed in a commercial product, or if you work primarily outside the Chinese social platform ecosystem. Before installing, resolve the licence question: the repository reports NOASSERTION while the README badge reads MIT-NC, and those two statements cannot both govern the same code. Then confirm that the 2.8 features you actually need are present in the installer you download, because the README itself warns that installer and public source may differ.

Official sources

1. Issues
2. Jamailar/Beav on GitHub
3. Project website
4. README
5. Releases