ยอดนิยม: อัปเดตข้อมูลเมื่อ 21 พฤศจิกายน 2569

หมวดหมู่

อินเทอร์เน็ต1,422
การสื่อสาร329
เสียงและวิดีโอ938
กราฟิกและการออกแบบ566
ออฟฟิศและเพิ่มประสิทธิภาพ1,185
เครื่องมือนักพัฒนา3,290
เครื่องมือ AI610
ความปลอดภัยและความเป็นส่วนตัว408
เครื่องมือระบบ3,201
เกม545
การศึกษาและวิทยาศาสตร์242
เครือข่ายและไอที294

ดาวน์โหลด LobeHub 2.3.4

An open-source, modern-design AI chat framework. Supports Multi AI Providers (OpenAI / Claude 4 / Gemini / Ollama / DeepSeek / Qwen), Knowledge Base (file upload / knowledge management / RAG), Multi-Modals (Plugins/Artifacts) and Thinking.

LobeHub ผู้ช่วย AI และ LLM #16 ผู้ช่วย AI และ LLM
นี่คือเวอร์ชันล่าสุด · วันที่เผยแพร่:

ไฟล์ตัวติดตั้ง 2.3.4

สถาปัตยกรรมประเภทขอบเขตขนาดไฟล์ดาวน์โหลดSHA256
64 บิต EXE (NSIS) ผู้ใช้ปัจจุบัน — github.com VirusTotal
64 บิต EXE (NSIS) ผู้ใช้ทุกคน — github.com VirusTotal

ลิงก์และค่า SHA256 มาจากแมนิเฟสต์ชุมชนของ Windows Package Manager โปรดเทียบค่าแฮชของไฟล์ที่ดาวน์โหลดก่อนเปิดใช้งาน

บันทึกประจำรุ่น

From PR #23286: https://github.com/lobehub/lobehub/pull/23286
🚀 LobeHub Release (20261118)
Release Date: November 18, 2026
Since v2.3.3: 291 commits · 291 merged PRs · 13 contributors
This weekly release turns Channels into a place where people and Agents work side by side, lets MCP connectors run on your own devices, and makes coding sessions easier to steer while they run.

✨ Highlights
- Agents in Channels — Agents can message each other or a person for their requester, answer in place or in a Thread, be muted until @mentioned, and wait for an offline device instead of failing the reply. Agents @mentioned together now gather in one Thread and divide the work. (#23139, #23244, #22911, #23164, #23060, #23103, #23222)
- MCP connectors on your devices — Run stdio and local MCP connectors on lh connect devices, route HTTP MCP servers through a device, subscribe to webhook events, and connect multiple accounts per OAuth connector. (#23047, #23056, #23067, #22991, #22988)
- Webhooks and automations — Manage webhooks and automations from the CLI and Agent tools, and deliver their events to Channels. (#23266)
- Better Coding Experience lab — Queue or steer follow-ups while an Agent runs, pick a branch for a new topic, switch or fork a worktree when sending, and link several pull requests to one topic, with the reason each was linked. Worktrees the app created are cleaned up when their topic is archived or trashed. (#23171, #23186, #23175, #23173, #23185, #23248, #23241)
- Persistent Sandbox — Choose a Persistent Sandbox from the execution environment; new conversations keep each Cloud Sandbox's last folder and repository. (#23126, #23122, #23212)
- DingTalk and WeCom bots — Connect DingTalk as an IM bot channel, and run the WeCom AI bot over its long connection with a guided setup. (#22336, #22326, #23089)
- Voice and music generation — Added ElevenLabs, Fish Audio, Doubao Speech and Gemini text-to-speech, plus provider-agnostic music generation and audio file preview. (#22168, #22951, #22977, #23000, #22918)
- A simpler composer — One fixed Send/Stop button, with the send-key choice moved to settings. Sending while files upload queues the message. Pasted and uploaded images show as animated cards, duplicate pastes are skipped, and quoted selections are kept in the message. (#23261, #23095, #23098, #23080, #23096, #23111)

🏗️ Core Agent & Architecture
Agent runtime
- Agents now park on a long-running shell command instead of polling it, and stop gracefully after repeated unknown tool names. (#22979, #23024)
- A running sub-agent can be steered with a follow-up or cancelled. (#22377, #22496)
- Agents can be woken from a generic webhook. (#23128)
- Runtime status reads no longer load the full state blob, tool-set copies stay out of per-step state, and inline steps read the topic once. (#23210, #23170, #23144)
- Manual compaction uses the topic-pinned model, and stale tool-result trims persist across warm turns. (#22697, #22981)
- Deleted Agent documents are no longer sent from topic history. (#23235)
- Pressing Stop no longer loses the messages of the step it interrupts. (#23269)
Goals and tasks
- Goals support marks between runs and evaluator-scored objectives, and a finished goal's report can be shared. (#23108, #22752)
- Search the Goals page and find goals from the command menu. (#23206)
- Find in view on the Tasks page with Mod+F; moving an unfiled task tree into a project moves its subtasks too. (#22960, #23214)
Memory and knowledge
- Memory extraction retries transient model outages, stays within the context window, and falls back from retired models. (#22894, #23088)
- Knowledge items support per-item access lists, enforced on content reads. (#22905)
- Uploaded-file parsing and chunking now run isolated from the main server. (#23117, #23260)

📱 Channels & Integrations
- Admins can make a group Channel public for the workspace to join; project Channels open to the project with project Agents first. (#23110, #23107)
- People can pin a conversation, edit their own messages, and move a long direct reply into a Thread. In a group, an Agent opens a long reply's Thread under the request and acknowledges inside it. (#23161, #23155, #23074, #23268)
- Thread previews appear under root messages and in the Channel list. (#22914, #23204)
- A message you send no longer counts as unread for you. (#23263)
- QQ group @mentions now reach LobeHub and get replies in the window. (#22921)
- A deferred bot message no longer replays itself in a loop, and WeCom keeps replies that arrive during a gateway restart. (#23147, #23092)

🖥️ Desktop & User Experience
- Desktop: Export a diagnostics bundle, write structured logs with a query CLI, and enroll the machine as a private workspace device. (#22966, #22873, #22826)
- Desktop: Reopen closed tabs with Cmd/Ctrl+Shift+T and switch tabs with Cmd+1–9; Recent Pages shows each page's own title. (#23224, #23223)
- Desktop: Staged Core OTA updates are now shown instead of reporting "up to date". (#23018)
- Sidebar: Drag to reorder team and personal pins, hide agent groups, and see task status in compact Favorites rows. Clipped topic titles scroll on hover, and double-clicking a topic no longer opens a new tab. (#23048, #23072, #23203, #23062, #23076)
- Working sidebar: The Overview shows Claude Code background tasks. (#23198)
- Projects: Folder card view, unified project environments with Settings → Environments, and a dedicated main agent for each new project. (#22860, #23150, #23129)
- Skills: Library Skills are opt-in per Agent, and the connector and Skill catalogs have a new layout. (#23188, #23187)
- Agent Profile: Simpler navigation and instruction editing. (#23275)
- Agent Share: Visitors can use the shared Agent's execution environment. (#23240)
- Subscription: Usage and Balance are merged into one settings page, and the Basic plan is available. Personal settings now list Subscription Accounts, split into Workspace and Private tabs. (#22985, #23090, #23270, #23250, #23242, #23245)
- Settings: Data export runs in the background and notifies you when the file is ready. (#23125)
- Member and Agent hover cards now appear across the app. (#22970, #23054)
- Long workflows stay responsive when opened and while streaming. (#23176)
- Answered question cards clear correctly, and avatars crop non-square images instead of stretching them. (#23063, #23064)

🔧 Tooling & CLI
- lh agent list --tag, lh agent create --tag, agent transfer and move-group commands. (#23016, #23012, #22923)
- lh acceptance delete removes an acceptance shell. (#23081)
- Library uploads from the CLI recover with same-account credentials. (#23259)
- Antigravity is supported as a local agent, and DeepSeek Harness models are discovered over ACP. (#22008, #22542)
- Follow-ups now steer native Codex runs on connected devices, and Codex warnings are no longer saved as tool calls. A model's weekly quota shows only when that model is selected. (#23264, #23252, #23071)
- A fixed private workspace Agent's owner can switch it to their private device. (#23230)
- Added the GPT-6.1 Terra model card and xhigh thinking for Gemini 4 Hydrogen. (#22805, #22933)

🔒 Security & Reliability
- Security: Server-side HTTP MCP requests go through the SSRF-safe fetch. (#23050)
- Security: Uploads now require reservation-backed upload sessions. (#20524)
- Security: The GitHub credential is no longer sent to devices for heterogeneous agents. (#23225)
- Reliability: Expected error conditions return 4xx or fall back instead of surfacing as 500. (#23058)
- Reliability: Refused async tasks fail cleanly, with JWT expiry and image type guards. (#23135)

⚙️ Upgrade Notes
- This release includes database migrations for contacts, inbound webhooks, connector visibility, Channel Thread locks, acceptance review rounds and artboard tables. (#23045, #22990, #22996, #23143, #22833, #22772) Self-hosted deployments apply them on startup; back up the database before upgrading.
- The im_layout rail layout and the legacy-plugin CLI create/install subcommands are removed. (#23084, #22372)

👥 Contributors
Huge thanks to 13 contributors who shipped 291 merged PRs this cycle.
@smbslt3 · @Coooolfan · @AleFlyX · @Max-Reisinger · @arvinxx · @tjx666 · @rdmclin2 · @ImSingee · @Innei · @AmAzing129 · @MarioJames · @ONLY-yours
Plus @lobehub-coder for agent-authored changes.

Full Changelog: v2.3.3...release/weekly-20261118
🤖 Generated with Claude Code

https://github.com/lobehub/lobehub/releases/tag/v2.3.4

ยอดนิยมใน ผู้ช่วย AI และ LLM

ดูทั้งหมด
  1. Claude Codeผู้ช่วย AI และ LLM · Anthropic PBC
  2. Codex CLIผู้ช่วย AI และ LLM · OpenAI, Inc.
  3. ChatGPT by lencxผู้ช่วย AI และ LLM · lencx
  4. GPT4Allผู้ช่วย AI และ LLM · Nomic
  5. oh-my-piผู้ช่วย AI และ LLM · Can Bölük
  6. Janผู้ช่วย AI และ LLM · Jan
  7. CC Switchผู้ช่วย AI และ LLM · farion1231
  8. T3 Codeผู้ช่วย AI และ LLM · T3 Tools
  9. WSL Managerผู้ช่วย AI และ LLM · Bostrot
  10. OpenWorkผู้ช่วย AI และ LLM · differentai

แท็ก

แชร์