WeKnora 0.9.0 Puts AI Agents on a Schedule—and Keeps a Record of Every Job They Do

WeKnora v0.9.0 is out. It is the open-source knowledge base framework from the Tencent WeChat team, with code on GitHub: https://github.com/Tencent/WeKnora . Two updates are the most practical this time: multi-agent collaboration and scheduled tasks.

The official description is straightforward: complex tasks are split into independent subtasks, assigned to agents, and the results are aggregated; tasks can be set to run once or repeatedly, with execution history and results available for review. For ordinary users, this can be understood as the AI in the knowledge base no longer just answering a single question, but being able to take on a job and run it on its own when the time comes.

Some commenters mentioned in replies that the real difficulty of enterprise knowledge bases is not searching documents out, but bringing sources back with answers, permissions changing with personnel, and old content expiring promptly. That judgment is closer to the problem itself than adding another chat entry point. This time WeKnora exposes execution history, associated conversations, and tool-call outputs, which at least gives scheduled tasks somewhere to be reviewed after they finish.

There are two details in the official replies: execution history shows not only the final result, but also lets you open the conversation associated with a particular run to see tool calls and outputs; runs of the same automation do not overlap—if the previous round is still running, the next round is skipped by default, or you can choose to queue one pending run to avoid pileup. For scheduled inspection-type tasks, this is more practical than simply piling on more agents. Some commenters said that scheduled tasks plus execution history are suitable for routine checks; previously many tools left no trace after running, so when problems occurred there was no way to review them.

v0.9.0 also adds execution permissions: read-only, confirmation before changes, and automatic writes, with scoped authorization for sessions and automations. The sandbox supports optional per-user reuse; background shell tasks can be managed; skill publishing is controlled. Parsing, sync, and retrieval have also improved: configurable image pipelines, OCR preprocessing, deeper Feishu/Lark sync, Confluence PAT, and retrieval improvements. WeKnora assistant can answer questions about version features and configuration, and create agents, scheduled tasks, and manage knowledge on the current page, executing under the current user's permissions.

For readers with a development background, several integration points are worth a look: the built-in MCP Server can publish the knowledge base to MCP clients such as Cursor and Claude; the CLI weknora outputs a stable JSON envelope by default, suitable for scripts and agent calls; Docker Compose, Helm, Lite single binary, and desktop app all have deployment paths. Ordinary users are more likely to encounter it through IM entry points such as WeCom, Feishu, Slack, and Telegram, or embedded components.

The hard problems of open-source knowledge bases will not disappear because of one version. How to maintain retrieval quality after documents get messy, how to implement permissions that follow people, and how to trigger expiration of old content—these still depend on actual deployment. On GitHub, some commenters asked about the website being a bit broken, and others asked how scheduled tasks handle timeout overlap; the official answer was no overlap and queue one. The questions are all very specific, more useful than launch-event-style praise.

WeKnora uses the MIT license, and both code and documentation are on GitHub. If you want to try it, Docker Compose is not too heavy to get started: clone the repository, copy .env, docker compose pull, up -d, then open localhost. The documentation site has installation, configuration, API, and release notes. Addresses: https://github.com/Tencent/WeKnora and https://weknora.weixin.qq.com/docs/

Putting AI on a schedule is not rare; what is hard is being able to check, after the schedule is set, what it did, why it did it, and where an error started. v0.9.0 puts execution records and tool calls on the table. This direction is more worth following than another chat entry point.