Tencent's Open-Source AI Agent Stack: WeKnora, BrowserSkill, LoopForge

Over the past year Tencent has open-sourced three AI infrastructure tools that address the practical gaps preventing agents from being usable in production: knowledge supply, browser entry, and coding discipline. None of them are foundation models, and none chase benchmark scores.

WeKnora: One Knowledge Base, Three Consumption Modes

Project: Tencent/WeKnora | Stars: 31,343 | Forks: 4,188 | Language: Go | License: MIT | Latest: v0.8.2 (2026-09-24, near-monthly releases) https://github.com/Tencent/WeKnora WeKnora avoids being "yet another RAG" by exposing the same document corpus through three distinct interfaces:

RAG Q&A — answers with cited sources for traceability.

Agent Reasoning — multi-step tool use to complete tasks (e.g., 20 tool calls in a demo).

Auto Wiki — transforms scattered documents into interlinked pages and a knowledge graph.

Originating from the WeChat ecosystem (official domain under weixin.qq.com), WeKnora integrates with 10+ IM platforms (Enterprise WeChat, Feishu, Slack, Telegram, etc.) and embeds a web Q&A widget. It auto-syncs from Feishu Docs, Confluence, Notion, Yuque, and other sources.

Technical stack: Go backend + Vue frontend. Eight pluggable vector stores (pgvector, Elasticsearch, Milvus, Qdrant, etc.). Integrates 27 model vendors. Retrieval uses hybrid search with reranking; GraphRAG is optional. Official benchmarks show support for 40,000-document knowledge bases.

v0.8.2 adds two notable features: a built-in MCP Server (one /mcp endpoint per workspace) that publishes the knowledge base as an external brain for tools like Cursor and Claude Code, and a "local browser" control capability that ties into BrowserSkill (see below).

Deployment is three commands:

git clone https://github.com/Tencent/WeKnora.git && cd WeKnora cp .env.example .env # edit model and storage config docker compose up -d

The documentation explicitly warns: production must run on a private network, never exposed to the public internet. Full functionality means a large attack surface — do not run unprotected.

BrowserSkill: The Agent Uses Your Actual Browser

Project: Tencent/BrowserSkill | Stars: 7,928 | Forks: 558 | Language: Rust + TypeScript | License: MIT | Latest: cli-v0.3.2 (2026-09-30, frequent updates) https://github.com/Tencent/BrowserSkill Most browser automation approaches fail because cloud browsers lack the user's login state, cannot reach internal networks, and require re-authentication. BrowserSkill inverts this: it lets the agent drive the browser you are already using, inheriting your logged-in sessions and intranet access.

The runtime consists of three pieces: a Rust CLI ( bsk) with a resident daemon, plus a browser extension. Once installed, any terminal-capable agent (Cursor, Claude Code, Codex, OpenClaw, and ~10 others) gains a set of browser tools.

Key design choices:

Tasks run in a dedicated Agent Window — visible and stoppable at any time.

Borrowing a tab triggers a confirmation prompt; the tab is returned after use.

Steps only a human can perform (login, CAPTCHA) are explicitly handed back — human-in-the-loop is a built-in mechanism, not an afterthought.

Trade-offs are stated plainly: the Agent Window is not a security sandbox; the agent operates with your identity. The project repeatedly advises delegating only to trusted agents and tasks. Privacy posture is clean: no mandatory cloud service, no product telemetry, the extension does not independently call AI vendors.

Installation in three steps:

curl -fsSL https://raw.githubusercontent.com/Tencent/BrowserSkill/main/install.sh | sh # install browser extension, enable local connection bsk install-skill # equip your agent with browser skills

Near 8,000 stars in three months with same-day releases indicates strong market validation for this "browser connection layer" approach.

The three projects form a loop: WeKnora v0.8.2's "local browser" feature is officially implemented on top of BrowserSkill.

LoopForge: Giving Coding Agents a Structured Workflow

Project: Tencent/LoopForge | Stars: 48 | Forks: 14 | Language: Python | License: MIT | Latest: v0.2.2 (2026-09-16) https://github.com/Tencent/LoopForge LoopForge does not build a coding agent; it governs how coding agents work. It equips CodeBuddy, Codex, Cursor, and Claude Code with a shared multi-stage workflow: understand requirements & confirm boundaries → design → implement → independent review → test & deliver.

It targets real pain points: starting before requirements are understood, scope creep, self-review without independence, long tasks losing context on disconnect.

Two practical designs stand out:

Delivery artifacts — at task end, confirmed requirements, technical design, independent review findings, and test results are archived into an artifacts directory, providing an auditable trail beyond a mere diff.

Breakpoint resume — state is persisted; a single command continues after interruption.

A local observability dashboard surfaces per-session token consumption, cost curves, and tool failure rates — data usually locked inside each agent's black box.

The project is candid about its maturity: created late July 2026, 48 stars, no topics tagged yet, v0.2.2. The README explicitly states that one-off small changes should just use a coding agent directly; LoopForge is positioned for medium-to-large tasks and team engineering that requires traceability.

Usage:

npx -y loopforge-cli@latest install codebuddy # also supports codex / cursor / claude # start in session: /start-devflow "task description" # resume: /resume-devflow {task_slug}

Official warning: some host configurations enable auto-execution mode; use only in trusted projects.

Conclusion

Viewed together, Tencent's positioning is consistent: they avoid the model layer and focus entirely on the prerequisites for making agents actually work — knowledge for context, browser for entry, workflow for discipline.

The posture is also consistent: all MIT, no cloud lock-in, no mandatory telemetry, human-in-the-loop encoded as mechanism rather than slogan.

As model capabilities become commoditized, the value of this connective tissue will only grow. WeKnora has proven the model, BrowserSkill is validating it now, and LoopForge has just started.