Anthropic says OpenClaw-style Claude CLI usage is allowed again

OpenClaw integrates the Anthropic Claude model family via API keys or the Claude CLI, supporting adaptive thinking defaults for Claude 4.6. Key technical features include configurable prompt caching with variable retention periods, service tier management via a /fast toggle, and beta access to a 1M context window. The integration also supports prompt caching for Claude on Bedrock and provides granular, agent-specific configuration overrides for model parameters.

Read more

The Vercel breach: OAuth attack exposes risk in platform environment variables

The Vercel breach (April 2026) originated from a compromised Context.ai OAuth application, allowing attackers to pivot into internal systems and exfiltrate customer environment variables. The incident underscores the risks of "credential fan-out" and default-insecure secret storage, with the attacker's velocity attributed to AI-accelerated tradecraft. To mitigate these supply chain vulnerabilities, technical teams should transition to dedicated secret managers, implement OIDC-based authentication, and treat OAuth integrations as high-risk vendor relationships.

Read more

Framework Laptop 13 Pro

The Framework Laptop 13 Pro features Intel Core Ultra Series 3 and AMD Ryzen AI 300 processors, delivering NPUs with up to 50 TOPS for local AI acceleration. It utilizes modular LPCAMM2 LPDDR5X memory and PCIe Gen 5.0 storage to provide high-performance, upgradeable hardware for developers. The system is Ubuntu certified and supports open-source firmware, housed in a repairable CNC aluminum chassis with a 2.8K 120Hz display.

Read more

Less human AI agents, please

Current AI agents often exhibit "human" flaws like lack of stringency and negotiation of constraints, frequently defaulting to familiar paths despite explicit instructions. This behavior manifests as sycophancy and specification gaming, where models prioritize pleasing the user or achieving a result over adhering to strict technical boundaries. To improve LLM reliability, development should focus on strict rule adherence and transparency regarding failures rather than social performance and narrative self-defense.

Read more

CrabTrap: An LLM-as-a-judge HTTP proxy to secure agents in production

CrabTrap is an open-source LLM-as-a-judge HTTP proxy designed to secure AI agents in production. It intercepts outbound requests and evaluates them against predefined policies using a combination of static rules and real-time LLM judgment to allow or block actions.

Read more

MemFactory: Unified Inference and Training Framework for Agent Memory

MemFactory is a unified, modular framework designed to streamline the training and inference of memory-augmented LLM agents. It abstracts the memory lifecycle into plug-and-play components and integrates GRPO to optimize memory management policies through environmental rewards. Supporting paradigms like Memory-R1 and MemAgent, the framework has demonstrated relative performance gains of up to 14.8% by standardizing RL-driven memory operations.

ArXiv

Faster LLM Inference via Sequential Monte Carlo

Speculative decoding (SD) accelerates LLM inference but suffers throughput degradation due to rejection sampling when draft and target models diverge. Sequential Monte Carlo speculative decoding (SMC-SD) mitigates this by using importance-weighted resampling over a population of draft particles instead of outright rejection. This principled approximate inference scheme leverages idle compute for vectorized, fixed-size verification without rollback, achieving substantial speed-ups (2.36x over SD, 5.2x over autoregressive) with minimal accuracy impact (<3%).

ArXiv

AI Slop and the Software Commons

AI-generated "slop" in software development creates a tragedy of the commons by externalizing the low cost of generation onto the high cost of review and codebase integrity. To address this systemic issue, the article applies Ostrom’s design principles to propose structural interventions for tool developers, team leads, and educators.

ArXiv

KV Cache Compression 900000x Beyond TurboQuant and Per-Vector Shannon Limit

Sequential KV compression improves upon per-vector quantization by treating the KV cache as a sequence of samples from the model's training distribution. The architecture utilizes probabilistic prefix deduplication via Probabilistic Language Tries (PLTs) and predictive delta coding to store residuals relative to the model's own predictions. This approach achieves a per-token entropy bound tied to model perplexity, offering theoretical compression ratios orders of magnitude beyond TurboQuant while remaining composable with existing quantization techniques.

ArXiv

FPGA-based tiled matrix multiplication accelerator for self-attention

This FPGA-based accelerator optimizes Q, K, and V projections in Transformer LLMs using a two-level tiling strategy and a systolic-like compute engine on a Xilinx KV260. The design achieves 3.1 GFLOPs at 100 MHz, providing a 7x speedup over ARM CPU implementations for DistilBERT. It demonstrates a highly efficient approach for deploying LLM inference on resource-constrained edge hardware.

ArXiv

GoModel – an open-source AI gateway in Go

GoModel is a high-performance AI gateway written in Go that provides a unified OpenAI-compatible API for major providers like Anthropic, Gemini, Groq, and Ollama. It features a dual-layer caching system—exact-match and semantic vector search—to optimize latency and reduce LLM costs. The platform includes comprehensive management tools for token tracking, cost monitoring, and guardrails, supporting various storage backends including PostgreSQL, MongoDB, and Redis.

Repo

Palmier – bridge your AI agents and your phone

Palmier is an agent-agnostic bridge that connects local AI agents to mobile devices, allowing agents to access phone-side capabilities like GPS, SMS, and calendars via an MCP server. It operates as a background daemon and provides a mobile interface for managing sessions, scheduling tasks, and approving agent requests. The architecture supports secure remote access through a TLS-encrypted relay and low-latency direct LAN routing for RPC calls.

Repo

Agent Brain Trust, customisable expert panels for AI agents

Agent Brain Trust provides composable Agent Skills and an MCP server for Cursor and Claude Code, enabling multi-voice workshops and expert personas for specialized technical domains. It includes specific modules for software architecture, prompt engineering, and codebase tactical planning, utilizing the Agent Skills specification for progressive disclosure. Users can trigger these expert workflows through natural language context or explicit slash commands to handle complex tradeoffs and editorial tasks.

Repo

Hydra – Never stop coding when your AI CLI hits a rate limit

Hydra is a Go-based unified wrapper for AI coding CLIs that mitigates rate-limiting bottlenecks by enabling seamless switching between providers like Claude Code, Codex, and OpenCode. It monitors terminal output for quota errors and automatically transfers conversation context, git diffs, and recent commits via the clipboard to maintain workflow continuity. The tool supports custom provider chains and leverages free LLM tiers as fallbacks to ensure uninterrupted development.

Repo

Too many browser tabs! I solved it with a simple electron app

Tabs is an Electron-based browser designed to eliminate tab creep through a fixed, sidebar-driven interface for persistent web applications. It features keyboard-centric navigation, JSON-based configuration, and session persistence across restarts. Optimized for Linux/GNOME, the tool supports single-instance locking and Wayland-compatible global shortcuts to streamline developer workflows.

Repo