Easel

Easel is an open-source workbench that runs a social-media account end to end, built on the OpenClaw agent runtime and shipped as a Python command line plus a React console on port 7860. Its capabilities are organised as five consecutive stages — discover, plan, produce, publish, attribute — and they are directories rather than menu entries: each skill is a SKILL.md of under two hundred lines, with reference documents loaded only when needed and runnable scripts beside them, so cards, voice-overs, subtitles, edits, short dramas and paper explainers come out as files written into an outputs/ project folder. An account profile of six dimensions — positioning, style, audience, platforms, preferences and red lines, long-term memory — is prefixed onto every turn, which is what makes the output drift toward a real account instead of a generic model voice. Publishing covers seven platforms, each with a guard before the real send, and the account data comes back through the same skill system, so the next round starts from what the last one measured.

Who built itZJU-REAL is a laboratory account rather than a person: its README carries Zhejiang University, Peking University, REAL Lab and OpenDCAI Lab marks, and frames the project as research moved into real social-media work. Its 210 commits are spread over thirteen author names. A hundred and forty-three of them arrive from an address at xiaohongshu.com and are not linked to a GitHub account; the account lidingm, whose commits are signed Dingming Li, has nineteen more; mengyuyuan has twenty-three. Sixty-five of the 210 carry a linked account, and twelve carry a co-author trailer, seven of which name a Claude model — six say only “Claude”, one names Opus 4.8 with a million-token context. GitHub lists ten contributors, and the repository sits at 2,633 stars and 402 forks.

1. 01

   Thirty-four days, four releases, and a repository that is mostly video

   The repository was created on 2026-08-28 at 02:22 UTC and its two hundred and tenth commit landed on 2026-10-01, so the whole record covers thirty-four days: fifty commits in the three remaining days of August, a hundred and fifty-nine in September, one on the first of October. Four releases came out of that stretch — v0.1.0 on 2026-08-31, v0.1.1 on 2026-09-14, v0.2.0 on 2026-09-17 with video production built in and three tiers of transcription, and v0.2.1 on 2026-09-24 covering the settings panel, faster conversation over the gateway and security hardening. Around them sit 2,633 stars, 402 forks, 14 watchers, 24 open issues and an Apache-2.0 licence. The tree holds 988 files and 330 MB, and almost none of that is code: assets/readme/ is 59 files and 239 MB — nine complete films, the largest 57,580,062 bytes, with six-second previews and thumbnails beside them — and web/static/ adds another 206 MB of the same material for the product page. The skills are 716 files and 12.9 MB, the shared scripts 55 files and 1 MB, the React console 46 files and 519 KB, the tests 110 KB, the Python command line 44 KB, the documentation 52 KB. Sixty-five of the 210 commits carry a linked GitHub account and twelve carry a co-author trailer; seven of those trailers name a Claude model, six of them saying only “Claude”.

2. 02

   Seven platforms, two ways in, and a dispatcher above both

   Seven platforms support login, adaptation and publishing: Xiaohongshu, Douyin, Kuaishou, Zhihu, Bilibili, WeChat Channels and WeChat Official Accounts. A contributor proposing the overseas publishers states the domestic layer plainly — browser automation or the platform command line. Xiaohongshu is a 47,674-byte Playwright script. One pull request adds a cross-process lock in the browser profile, msvcrt on Windows and fcntl elsewhere, because a background login probe and a foreground publish could both launch Chromium against one directory and the process died with exit 21; the same change prefers an anti-detection kernel under ~/.cloakbrowser, after headless mode kept meeting the platform’s 300012 challenge. Douyin is 76,086 bytes, the largest script here, with a headed login whose window must not be closed early and a keep-open mode forbidden from re-opening the browser context, since a second lock on one profile crashes. Bilibili calls biliup, a third-party uploader. WeChat Official Accounts is the one that talks to a platform API: images in the body are uploaded through the material endpoint and replaced with CDN links before the draft is created, because the platform will not fetch a local path. Above them publish_dispatch.py holds the platform registry and routes an adapted piece to the right publisher.

3. 03

   The gate before the real send, and what to do when the result is unknown

   Every script that puts text on a public platform must call the outbound content guard before the real send. It is a 21,848-byte module with two levels: a blocking finding is real sensitive information — API keys, internal URLs and domains, proxy addresses, internal paths, environment names with their real values — and it stops the publish with exit code 7, failing closed; a warning is AI-flavoured wording or a model name, which in a paper explainer may well be legitimate, so it only warns. Dry runs never block, and overriding a hard block takes an explicit flag. Publishing is dry-run until an explicit switch is added. The harder decision is what to do when the outcome is unknown, and its author wrote the first version wrong and corrected it in public. Only an explicit 400, 401, 403 or 422 counts as a definite rejection. A 5xx, a timeout, a dropped connection, or a 2xx with an unusable body is not a failure, and the script does not resend: it queries the same request identifier it had sent as an idempotency key. If that finds the post it keeps tracking; if not, the attempt is reported as unresolved, written into the calendar as pending with the identifier in the note, and kept out of the publish log so nobody later reads it as a failure and posts twice. The skill states the rule: after an unresolved result, do not run the command again — use the status command to check.

4. 04

   Reading the numbers back is its own set of scripts, with its own bugs

   The attribute stage is where the loop closes, and it is heavy: a 48,135-byte account statistics tool, a 41,251-byte readback module for the per-platform shapes, a 33,769-byte WeChat Official Accounts reader, a publish analytics skill with a follower-log schema, a performance review with a 7,640-byte benchmarks file, and a post scorer with a 4,084-byte script and a 4,888-byte criteria document. The bugs are instructive. One pull request reports that Xiaohongshu statistics were counting drafts: the creator centre opens on an all-notes tab, the script read that list, and untitled drafts and half-finished notes counted as published work. The fix clicks through to the published tab first, then filters out draft markers, empty titles and editor URLs. The same pull request protects Douyin’s readback, recording a plain instruction when a headed login waits for a scan and forbidding the keep-open path from re-opening the context to collect numbers. The loop reads plays, interactions, comments and performance and deposits the effective structures back into the account profile. The mechanism is a manifest contract: every layer records the paths it produced and a one-line summary into one hidden metadata file per project, so a downstream step reads the upstream file by path instead of re-deriving it, and a failed step is recorded as failed so a run can be resumed.

5. 05

   Two validators for the skills, and four bugs that were not about the agent

   One script checks every skill directory — frontmatter, resource links, the output and publishing safety contracts. A second parses the Python commands named inside each SKILL.md and compares them with the scripts’ own argument definitions, to catch drift between prose and code. There are thirteen test files, a continuous integration workflow, and an offline self-test inside each publishing script. The failures in the bug list are plumbing, not judgement. Running the test suite overwrote the real configuration file with fixture values, so the primary model reference no longer matched the provider catalogue and every later conversation reported an unknown model; five places in the web server touched that file directly, bypassing the module that already had an isolation mechanism. The gateway address was hardcoded as 18789 in thirteen places across seven files while the dedicated profile listens on 37289, so the checker printed a failure while the gateway was healthy. On Windows, the runtime locator could land on an npm shim, and because such a file runs through the command interpreter a multi-line message was cut at its first newline: the agent received the eighty-seven-character profile prefix and nothing else. The fix resolves the shim’s sibling runtime and module entry point instead, and the same prompt then measured three hundred and fifty-five characters.

6. 06

   Outside contributors, a full chat group, and a warning about one platform

   Of the thirty issues and pull requests, most come from outside the laboratory. One contributor has eight, spread across the settings panel, the installer, the command line and the image pipeline. Another has four in a week: Xiaohongshu publishing, a statistics fix, images in WeChat Official Accounts drafts, and a security pass that tightened cross-origin access on the local port and blocked deletion of system directories. Two contributors filed two each, including a chat bug where pressing return halfway through typing pinyin sent the message, and a dark-mode toggle. One proposed feature opened an issue before the code and offered to be closed if the direction did not fit. The maintainers route support to a WeChat group — one issue is simply that the group passed two hundred members and could not be joined — and in one thread the main committer suggests contributors reach for a working agent to debug their own problems, since machines and configurations differ. The published roadmap has four items, and the community pull requests are already hitting the second, simpler installation. Finally there is the warning in the README itself, unusual in a launch document: automated publishing to Xiaohongshu may be detected by the platform and carries verification, rate-limiting or account risk, so preview and pre-publish checks with a human confirmation are recommended.