Beav: a local-first AI workspace for Xiaohongshu creators

What Beav actually solves for a Xiaohongshu operator

The README opens with the problem in operational terms: running an account means reviewing a large volume of material, tracking what is trending, and hunting for angles, every day. General-purpose AI tools, the README argues, are built for programmers or office work and do not carry a long-lived asset library for a creator. That is the gap Beav claims.

The intended user is narrow and specific. Someone who operates one or more Xiaohongshu accounts, or a brand account, and wants collection, comment capture, topic selection, drafting, image generation and talking-head video editing to share one storage layer. The README's own framing is that Beav is a 素材库 and 运营工作台 for self-media workers, not a general assistant.

If you are a developer looking for a scraping library, this is the wrong shape of product. Beav is a desktop application with a browser extension, an optional MCP plugin for host agents, and a local workspace per account or brand. The unit of value is the workspace, not the function call.

How the workspace, collection and AI layers connect

The repository is a pnpm monorepo. The top level holds desktop/, plugins/, Plugin/, scripts/, images/, and the docs (README.md, readme_en.md, CHANGELOG.md, ROADMAP.md, AGENTS.md), with package.json pinning Node to >=22 <23 and pnpm to 10.28.2. The TypeScript source lives under desktop/ and plugins/; the dependencies listed at the root are small and telling: diff, vscode-jsonrpc, and the vscode-languageserver-protocol and -types packages. That is the shape of a client that talks to a language-server-style backend over JSON-RPC rather than a web app with a REST API.

The data flow the README describes runs in one direction. A Chrome or Edge extension saves web pages, social content and comments into the local workspace. Subscribed creator pages refresh daily into the knowledge base. Automated research queries multiple social platforms, filters candidates, then reads the high-value ones in depth, and the results land in the knowledge base where later drafting and topic work can retrieve them. The daily 运营日报 pulls news and trending items with source links, and you can move an item from the report into a topic or straight into drafting.

Two details matter for anyone evaluating the architecture. First, the README states the workspace is local-first and that material, drafts and projects are organized around a local workspace. Second, it states models are selectable: the official AI or an OpenAI-compatible endpoint with your own Endpoint, API Key and model. So the AI layer is pluggable, but the storage layer is not described as a server you can point somewhere else.

The 2.8 release adds a media workshop. Imported talking-head video is transcribed word by word, then analyzed for semantic and pause-based redundancy so retakes, filler words and seams can be proposed for removal. The timeline supports multiple tracks, canvas transforms, keyframes, transitions, color grading, masks and audio adjustment, and exports SRT or WebVTT subtitles separately. The README is careful here: picture-in-picture placeholders represent material still to be added, and an animation plan is not a finished frame. You are expected to check the timeline, the playback and the exported file before using the result.

Installing Beav and getting a first workspace running

Beav is not installed from a package registry. The README points at a download page, and the repository's release assets carry the installers, extensions and signatures. The stated sequence is: install Beav, create a workspace for one account or brand, configure AI, then get the browser extension if you need collection.

The download page links are given in the README for the mainland China network case, and the latest release page is where the Chrome and Edge extension lives. There is no npm install step in the documented flow, and the repository's package.json is the build-side manifest rather than an end-user entry point.

Once Beav is running, the AI configuration step is the one that decides whether the rest works. In `设置 → AI` you either use the official AI or supply your own Endpoint, API Key and model. The README does not print the exact form fields beyond those three names, so treat the labels as the contract.

For CLI users, the README gives one command to bring the app up before connecting an external agent:

beav open

After that, the documented first real task is to start from the homepage topic entry or the daily report, and if you already have video, to open the media workshop for editing. Small tools such as the subtitle extractor, the image metadata cleaner and the account planning tool open from the homepage and request permission on demand.

If you want a host agent such as Codex Desktop or WorkBuddy to drive Beav over MCP, the README gives the install instruction as a prompt rather than a command:

/goal Read https://beav.ziz.hk/agent to install the Beav Creator plugin and set up a new task for me.

The README states that after installation you describe the task in the host agent, and that you still review, approve and edit results in the Beav UI. The browser UI is described as view-only in that arrangement, which is a meaningful constraint if you expected to keep working in the app while an agent runs.

Where Beav stops: licensing, platform gaps and unverified paths

The licence is the first hard boundary. The README badge reads MIT-NC, while the repository metadata reports NOASSERTION. Those two do not agree, and the practical reading is that the terms are non-commercial in the README's own framing. If you plan to use Beav inside a commercial content operation, read LICENSE in the repository root before you build a workflow on it. That is a factual gap in the metadata, not a legal opinion.

The second boundary is platform coverage inside the tools. The README states the image-to-Live-Photo tool currently supports macOS and Windows, and that Linux is not supported. It also states that Windows real-device testing and phone import are still pending acceptance, and that paired files must be kept together during transfer. So even on a supported OS, one of the small tools is explicitly not fully signed off.

The third boundary is the version skew the README admits to directly: the install package and the public source may update at different rates, and you should treat the entry points of the version you actually installed as authoritative. That is an unusually honest line, and it also means a bug you find in the repository may not exist in your build, or vice versa.

Finally, the online services are conditional. The README says creating a template and submitting an image generation are two separate operations, and that online generation depends on account permissions and model configuration. Automated social research and creator subscription both depend on the current data sources and configuration. None of these are described as guaranteed.

Beav versus assembling the same stack yourself

The obvious alternative is not a single competitor. It is a composition: a downloader for Xiaohongshu notes and comments, a note-taking or knowledge tool, a general chat assistant for drafting, and a separate editor for video. The README names Codex and WorkBuddy as the general-purpose tools it is positioning against, arguing they are built for programmers or general office work rather than for a creator's long-lived asset library.

The difference in approach is where the state lives. A composed stack keeps each artifact in the tool that produced it, and you are the integration layer: you copy a comment export into a notes app, paste a draft into a chat window, and re-upload images into an editor. Beav's claim is that collection, knowledge base, drafts, image templates and video projects sit in one local workspace tied to an account, so the AI can retrieve prior material without you re-feeding it.

That trade is real in both directions. A composed stack lets you swap any piece, script against it, and run it headless on a server. Beav's integration is the product, and the same integration is what makes it hard to automate: the documented interface is a desktop UI, a browser extension, and an MCP plugin for host agents, not a documented HTTP API. If your workflow needs to run unattended on a schedule, the composed stack is the more honest choice.

Maintenance cadence and what an upgrade costs you

The repository is not archived, and the last push was on 2026-09-10. The release history in the same window shows v2.8.0 on 2026-09-10, v2.7.20 on 2026-09-04 and v2.7.19 on 2026-09-03, so the project is moving on a roughly weekly cadence at the point of the latest release. The README also links a CHANGELOG.md and a ROADMAP.md in the repository root, and states that installers, extensions, update assets and signatures are all published through GitHub Releases.

The upgrade cost is mostly not code. Because the README warns that the install package and the public source can diverge, and because it tells you to defer to the entry points of your installed version, an upgrade can move a menu item or rename a workspace tool without a corresponding source change you can read first. Your CHANGELOG.md is the thing to read before upgrading, not the repository diff.

The second cost is model configuration. If you use your own OpenAI-compatible endpoint, an upgrade that changes how the AI settings are read is a config migration, and the README only names the three fields (Endpoint, API Key, model). Keep those values somewhere outside the app.

The third cost is data portability. The README describes a local-first workspace and mentions exporting finished video, subtitles and portable project files, but it does not document a bulk export or a rollback path for the workspace itself. The README does not document rollback. If a version misbehaves, plan on restoring your own backup of the workspace directory rather than on a built-in downgrade.

Editorial conclusion

Adopt Beav if you run one or more Xiaohongshu accounts and want collection, comment capture, a reusable asset library and AI drafting in a single local workspace instead of five browser tabs. Do not adopt it if you need a headless pipeline, a documented HTTP API, or a permissive commercial licence, because the repository badge reads MIT-NC and the README documents a desktop and browser-extension workflow rather than a server. Before committing, verify three things on your own machine: that the installer you download matches the release you expect, that your chosen model endpoint works under 设置 → AI, and whether the browser extension you need is actually attached to the latest Release, since the README notes the package and the public source may not move in step.

Frequently asked questions

Official sources

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