What is WeKnora?
WeKnora is Tencent's open-source framework for turning a team's documents into something people and agents can question. It offers three modes over the same knowledge bases: quick question answering with hybrid search and citations back to the source passage, an agent mode that plans multi-step work with skills, MCP tools, web search and sandboxed code, and a wiki mode that writes interlinked pages with a knowledge graph.
It parses more than ten formats, including PDF, Word, Excel, images and XMind, and syncs from Feishu, Confluence, GitLab, Notion, Yuque, DingTalk Docs, Seafile and RSS. Answers can be served in WeCom, Feishu, Slack and Telegram, through an embeddable widget, or to Cursor and Claude via a built-in MCP server. The backend is Go, with swappable models (29 built-in vendors, Ollama included), vector stores and object storage, and it deploys with Docker Compose, a Helm chart or a single Lite binary on SQLite.
Is WeKnora worth it?
WeKnora is one of the broadest self-hosted knowledge-base apps available, and it moves fast. The repository has about 37,400 GitHub stars, shipped v0.9.0 on 29 October 2026 after v0.8.3 on 14 October and v0.8.2 on 24 September, and was still taking fixes on 9 November.
Its strength is that one set of documents serves three jobs. Quick answers cite the exact passage, the agent can search, read documents, call MCP tools and run skills in Docker, E2B or Cube sandboxes, and wiki mode builds a browsable knowledge graph with revision history and rollback. Team features are further along than in most rivals: workspaces with four roles, per-resource ownership, a per-workspace audit log, scoped API keys and per-endpoint MCP tokens with rate limits.
The security record needs reading before deployment. GitHub lists 10 advisories for 2026, all from January to March: three critical, covering command injection through the MCP stdio configuration and an SQL-injection bypass in the database query tool, each leading to code execution, plus a high-severity tenant access-control bug. The README itself tells you to keep the service off the public internet. JWT_SECRET and SYSTEM_AES_KEY ship blank for you to fill in, the optional MinIO and Neo4j profiles keep their stock passwords unless changed, and since v0.8.2 sandbox commands run as root.
The catch: weight and focus. The standard Compose stack asks for 4 CPU cores and 8 GB of memory before any local model weights, because the document reader bundles LibreOffice and Playwright, and the issue tracker shows about 532 open issues. Many connectors and default model vendors (Feishu, DingTalk, Yuque, WeCom, Hunyuan, Qwen, Zhipu) serve Chinese workplaces; Slack, Notion and Confluence are there, but Google Drive and SharePoint are not on the README's list.
Our take: RAGFlow is the closer rival for raw document parsing and Dify for building general LLM apps; WeKnora earns its place when you want answers, an agent and a maintained wiki over the same files, with roles and audit logs, on your own servers.
Where it earns its place
An internal-tools team at a 500-person company on Feishu
Answer staff questions from the wiki and policy documents inside the chat app
Enable Feishu wiki auto-sync and the Feishu channel, and give each department its own workspace with Viewer roles. Cost is a server of at least 4 cores and 8 GB plus model API calls; citations let staff check the source paragraph.
A developer whose team writes code with Cursor or Claude
Let the coding agent look up internal design documents
Create a workspace MCP endpoint scoped to the engineering knowledge base, with its own token and rate limit, and add it to the agent's MCP config over Streamable HTTP.
A consultant who wants a private research library on a laptop
Search and summarise a few hundred PDFs without cloud services
Run the Lite single binary, which uses SQLite and an in-memory queue, with Ollama models. No published hardware figures for Lite, so expect speed to depend on the local model.
Before you commit
- Setup
- Clone the repository, copy .env.example to .env, fill in JWT_SECRET and a 32-byte SYSTEM_AES_KEY, then `docker compose up -d`; the web UI opens on port 80 with an onboarding guide. Profiles add Neo4j, MinIO or Langfuse.
- Integrations
- 29 model vendors including OpenAI, DeepSeek, Qwen, Zhipu, Hunyuan, Gemini, LiteLLM and Ollama; pgvector, Milvus, Qdrant, Weaviate, OpenSearch and others; Feishu, Confluence, GitLab, Notion, Yuque, DingTalk Docs, Seafile and RSS sync; WeCom, Feishu, Slack and Telegram channels; a built-in MCP server and a CLI.
- Data and privacy
- Self-hosted: documents, vectors and files stay on your servers, and text goes only to the model provider you configure. Credentials are encrypted with AES-256-GCM. No telemetry is described in the installation docs. The 2026 advisories include three critical code-execution flaws; most list patched versions from 0.2.10 to 0.3.2.
- Cost at scale
- The software is free; cost is servers (4 cores and 8 GB to start, more with Milvus, OpenSearch or Langfuse) plus model API tokens for answers, agent runs and wiki generation.
- Limits
- About 532 open issues; the Helm chart caps uploads at 50 MB by default; the desktop app has no published installer and must be built from source; upgrades include database migrations that the release notes ask you to read first.
What’s good
- RAG answers, agent and wiki mode share one set of knowledge bases
- Citations link back to the source passage
- Workspace RBAC with four roles, audit log and scoped API keys
- Built-in MCP server and chat channels for WeCom, Feishu, Slack and Telegram
What’s not
- Three critical advisories in 2026, each leading to code execution
- Standard stack needs 4 cores and 8 GB before local models
- Sandbox commands run as root since v0.8.2
- Connectors lean towards Chinese workplace tools
Best for
Engineering teams at companies on Feishu, WeCom or Confluence that want a self-hosted, permissioned knowledge base with an agent on top.
WeKnora alternatives
WeKnora: common questions
- Is WeKnora worth it?
- WeKnora packs RAG, an agent and a self-writing wiki into one self-hosted app with real access controls, but its 2026 security record and its tilt towards Chinese workplace tools mean it belongs on an internal network, chosen deliberately. We rate it 3.9 out of 5.
- How much does WeKnora cost?
- Free and open source (MIT, with third-party components under their own licences); you pay for servers and model APIs. Hosted options on WeChat Dialog Open Platform and Tencent Cloud Lighthouse; pricing for those is not published in the project's docs.
- What are the downsides of WeKnora?
- Three critical advisories in 2026, each leading to code execution; standard stack needs 4 cores and 8 GB before local models; sandbox commands run as root since v0.8.2; connectors lean towards Chinese workplace tools.
- Who is WeKnora best for?
- Engineering teams at companies on Feishu, WeCom or Confluence that want a self-hosted, permissioned knowledge base with an agent on top.
- How much does WeKnora cost for a team or in production?
- The software is free; cost is servers (4 cores and 8 GB to start, more with Milvus, OpenSearch or Langfuse) plus model API tokens for answers, agent runs and wiki generation.
- Is WeKnora safe for company data?
- Self-hosted: documents, vectors and files stay on your servers, and text goes only to the model provider you configure. Credentials are encrypted with AES-256-GCM. No telemetry is described in the installation docs. The 2026 advisories include three critical code-execution flaws; most list patched versions from 0.2.10 to 0.3.2.
- What are the best WeKnora alternatives?
- The closest alternatives we have reviewed are RAGFlow, Dify, MaxKB and AnythingLLM.
Answers checked Nov 13, 2026.