Claude Tag: @Claude on Slack for Multiplayer AI Collaboration
Anthropic announced Claude Tag, a new way for teams to work with Claude directly inside Slack. Instead of a one-on-one chat, Claude joins channels as a persistent team member. Tag @Claude with a task, and Claude breaks it down into stages, uses connected tools, and responds in a thread — while you work on other things.
Key Features
- Multiplayer: One Claude per channel interacts with everyone. Conversations persist — anyone can pick up where another left off.
- Learns over time: Claude builds context by following channel conversations and connected data sources. No need to re-explain from scratch.
- Ambient behavior (optional): Claude proactively flags relevant info from across channels and follows up on stale threads.
- Async and scheduled tasks: Set a task and walk away. Claude can schedule work over hours or days.
- DM support: Tag Claude in DMs for private, tool-access tasks.
Internal Usage
Anthropic reports that 65% of their product team's code is now created by their internal version of Claude Tag. Usage has spread beyond engineering to metrics, support tickets, and bug root-causing.
Deployment & Control
- System admins configure tools, data sources, and channel scopes per Claude identity. Memories stay isolated per channel scope.
- Token spend limits can be set per organization and per channel.
- An audit log tracks all Claude actions and who requested them.
- Works with Opus 4.8.
Availability
Available today in beta for Claude Enterprise and Team customers. Replaces the existing Claude in Slack app; admins can migrate within 30 days. Introductory launch credits provided.
Setup: link Claude Tag to Slack workspace → grant tool access → set spend cap → test in private channel.
📖 Read the full source: HN AI Agents
👀 See Also
XLI: Open-Source Python Library for Claude Code-Style Terminal UIs
Building a coding agent? Terminal UX is half the work. XLI is an open-source Python rendering engine that replicates Claude Code's streaming markdown, tool cards, inline approvals, and slash commands without hijacking your terminal scrollback.
JANG Quantization Method Improves MLX Performance for Large Models
A new quantization method called JANG enables running large models like MiniMax-M2.5 and Qwen 3.5 on Apple's MLX framework with significantly better performance than standard MLX quantization, achieving near-native speeds while maintaining accuracy comparable to higher-bit quantizations.
Fine-tuned Qwen3.5-2B with RAG-Engram architecture improves grounded answer accuracy from 50% to 93% at 8K context
A developer fine-tuned Qwen3.5-2B with a custom RAG-Engram architecture to address the 'lost in the middle' phenomenon, improving correct answers at 8K tokens from 50% to 93% on real-world queries. The system uses a two-level approach with static entity embeddings and dynamic chunk navigation.
Agent Browser Shield: Free OpenClaw Extension Blocks Prompt Injection & Dark Patterns
PixieBrix releases Agent Browser Shield, a free source-available browser extension for OpenClaw that blocks prompt injection, dark patterns, and context pollution while cutting token usage.