Tencent has released WeKnora 0.8.0 on GitHub, an open-source knowledge platform that turns documents into a queryable RAG system, a ReAct agent, and a self-maintaining wiki. The 0.8.0 drop is aimed at self-hosted enterprise stacks.
The headline additions are a skill sandbox runtime with session-persistent Docker, E2B, and Cube backends; a tenant skill catalog that installs skills from ClawHub, SkillHub, git, or zip; and cross-session long-term memory that stores profiles, preferences, facts, tasks, and interests. Wiki Mode continues as the agent-built markdown knowledge base with a knowledge graph; this release keeps the editing, revision-history, and rollback path that treats wiki pages as living documents rather than static RAG chunks.
Confirmed
- WeKnora 0.8.0 is published at github.com/Tencent/WeKnora under the Tencent organization.
- Three core modes: RAG quick Q&A; a ReAct agent that orchestrates retrieval, MCP tools, the skill catalog, sandboxes, and web search; and Wiki Mode for interlinked markdown knowledge bases with an interactive knowledge graph.
- Skill sandbox runtime: session-persistent Docker / E2B / Cube backends, per-tenant network policy, snapshots, live progress, and file browse/edit. The local host-process backend is removed; Docker is opt-in.
- Tenant skill catalog: install from ClawHub, SkillHub, git, or zip, with personal and workspace environment variables.
- Cross-session long-term memory: profile, preference, fact, task, and interest categories, with auto-extract-and-confirm plus a search_memory tool.
- Ingestion and integrations: in-process anydoc parsing for common office formats; connectors including Feishu, GitLab, Notion, Yuque, DingTalk Docs, and RSS; 20+ LLM providers via direct adapters and LiteLLM; IM channels including WeCom, Feishu, Slack, and Telegram; scoped API keys and 4-tier workspace RBAC.
- Fully self-hostable modular architecture.
Unknown
- No independent benchmarks compare WeKnora's RAG accuracy or agent success rates against LangChain, LlamaIndex, or commercial RAG platforms.
- Hardware requirements and cost profiles for running skill sandboxes (Docker, E2B, Cube) at production scale are not published.
- No timeline is given for enterprise support, SLAs, or a managed service.
- The changelog does not detail migration paths or compatibility guarantees for existing v0.7.x deployments upgrading to v0.8.0.
- No latency or throughput figures compare the in-process anydoc parser with external parsing services.
Our take
WeKnora 0.8.0 is Tencent treating agent skills as something that needs a real containment boundary—not a host process with hope. The sandbox + tenant catalog move the project from passive retrieval toward runnable knowledge, which matters for anyone wiring agents into internal docs. Without independent agent reliability numbers or published sandbox cost profiles, treat the 0.8.0 claims as directional until teams can prove sandbox reliability on their own workloads.
