WeKnora: WeChat's Enterprise Knowledge Framework Turned Agent Infrastructure

⬅️ Back to Tools

WeKnora: WeChat’s Enterprise Knowledge Framework Turned Agent Infrastructure

What it isAn enterprise knowledge platform that grew from RAG into agent infrastructure: pipeline, agent, wiki, memory, sandbox
PlatformSelf-hosted via Docker Compose or Helm; Go backend with web UI, CLI, MCP server, and IM channels
PriceFree, open source (MIT); you pay for hosting plus LLM calls
Linkgithub.com/Tencent/WeKnora · weknora.weixin.qq.com

The WeChat team open-sourced the knowledge stack they actually run, and at 23.5k stars with 3.3k forks it is one of the most watched agent-infra repos on GitHub. It started as RAG and kept growing: a ReAct agent, auto-generated wikis, long-term memory, and with v0.8.0 (September 3, 2026) a full skill plus sandbox runtime. I traced the README, the v0.8.0 changelog, and the wiki, memory, and sandbox docs end to end. I did not deploy a cluster, so read this as a verified map, not a field report.

  1. Layer one is a context pipeline that takes ingestion seriously. Thirteen formats including PDF, Word, Excel, PPT, images, XMind, and audio, plus auto-sync from Feishu, Notion, Yuque, GitLab, Tencent IMA, DingTalk Docs, and RSS. Office files parse in-process with anydoc, and every upload batch can override parser, chunking, VLM, and ASR settings through a process_config. Retrieval runs BM25 sparse plus dense plus GraphRAG with parent-child chunking, and chunks themselves are editable in the UI with revision history and automatic reindexing.
  2. Layer two is a ReAct agent that decides per turn what to reach for. Knowledge base search, wiki tools, Excel analysis, MCP services including OAuth ones mid-conversation, or web search across eleven providers. @Skill and @MCP mentions narrow the runtime for a single turn. The mental shift is the one that matters: the knowledge base stops facing users directly and becomes the agent’s long-term context layer.
  3. Layer three, new in v0.8.0, is the execution environment. Every chat session gets a persistent sandbox on Docker, E2B, or CubeSandbox backends, where the agent runs shell commands, reads and writes files, takes attachments from /workspace/input, and delivers results as artifacts in a per-message drawer. Skills install from ClawHub, SkillHub, GitHub or GitLab URLs, or plain zips, each with per-config network policy. The old local host-process backend was removed outright, and Docker is opt-in because a mounted docker.sock is host root. They say that part themselves, which I like.
  4. Memory is genuinely cross-session, not a context trick. Five typed stores (profile, preference, fact, task, interest) in explicit-only or auto-extract mode, capped at 200 items per identity by default. The restrained design choice stands out: system-inferred entries sit in pending until you confirm them, and unconfirmed guesses never reach the prompt. It also tracks documents you keep citing and lets that affinity condition future retrieval ranking.
  5. Auto-Wiki is the feature I would have dismissed as a gimmick, and the docs changed my mind. The agent distills raw documents into interlinked Markdown pages with a knowledge graph, six page types, folder hierarchy, and search. Ten wiki tools let the agent read, write, rename with backlink repair, and delete with dead-link cleanup. Pages get flagged with typed issues like mixed entities, contradictory facts, or out of date, and a built-in Wiki Fixer agent works the queue. Every edit carries version history with diffs and one-click rollback, capped at 50 auto and 200 total versions.
  6. Tencent even bridged this into coding agents. The official DeepSeek Harness plugin adds four read-only tools (weknora_search, weknora_read_document, weknora_ask, weknora_list_knowledge_bases), so a coding agent pulls company knowledge without ever seeing the underlying documents. The MCP server goes further with 29 tools published as a PyPI package. Private knowledge stops being a separate app and becomes a tool call.
  7. The enterprise bones are all there. Four-tier workspace RBAC with per-KB ownership and audit logs, capability-scoped API keys, 20-plus LLM providers plus LiteLLM gateway support, Langfuse tracing of every ReAct loop, IM serving through WeCom, Feishu, Slack, Telegram, and WeChat itself, an embed widget for external sites, and a Chrome extension for one-click capture. The product docs cover some 360 endpoints, which tells you this is meant to be integrated, not just tried.
  8. Honest limits. This is a Go backend plus frontend plus docreader plus vector stores, not a pip install; deployment means Docker Compose at minimum. Wiki generation, memory extraction, and reranking all bill LLM calls, so pick the focused extraction granularity before pointing it at 40k documents. The release cadence is brisk (0.6.2 to 0.8.0 in under three months), so pin versions. And my verification stopped at the docs; run the compose file before trusting any of this with real company data.

Install & first run 

git clone https://github.com/Tencent/WeKnora.git
cd WeKnora
cp .env.example .env   # edit as needed
docker compose pull && docker compose up -d

Visit http://localhost for the web UI, with the backend API on port 8080. Optional profiles add Neo4j for the knowledge graph, MinIO for object storage, or LangSmith-style tracing via Langfuse (--profile neo4j, minio, langfuse). Then the agent-first CLI takes over:

weknora kb list
weknora doc upload notes.md
weknora chat "summarise the design doc"

Every CLI command emits JSON by default for agent consumption. The README, the docs site in website-docs/, and the troubleshooting FAQ cover everything past this point, and they are better maintained than most commercial docs I have read.

Worth your time if: company knowledge is scattered across Feishu, Notion, and GitLab and you want agents that act on it inside a controlled sandbox, not just answer from it. Skip it if you need a personal notes search, that is a weekend project, not this.

Related TMFNK Content

Crepi il lupo! 🐺