Open Source Releases

OpenWork — Self-hosted collaborative AI coding environment built on opencode. If you want Claude Cowork vibes without shipping your codebase to Anthropic, this is the ticket. Team-based AI assistance with full data control.

ANE — Reverse-engineered Apple Neural Engine APIs that unlock training on Apple Silicon, not just inference. Previously ANE was inference-only; this cracks it open for on-device training workloads. Impressive reverse-engineering work, though stability on future macOS versions is anyone’s guess.

Creed Space — MCP server serving 16 tools for machine-readable value systems: persona selection, constitution merging, prompt generation, action adjudication. No auth, public endpoint. Essentially a constitutional AI toolkit you can plug into any agent. Useful if you’re building alignment guardrails and don’t want to roll your own.

django-pydantic-agent 0.4.3 — Settings-agnostic Django substrate for Pydantic-AI agents. Configuration-driven instantiation, no opinionated setup. If you’re running Django and want agents without fighting the framework, this saves the glue code.

django-ag-ui 0.26.2 — Django + Pydantic-AI + AG-UI integration. Async views, tool registry, audit logging for agent-driven UIs. The audit logging alone might justify it if you need traceability.

langguardx 0.4.0 — Security middleware for LangChain SQL agents. Adaptive multi-layer protection against injection and unauthorized data access. If you’re exposing SQL agents to users, you need something like this — preferably before the first penetration test.

Research Worth Reading

Kernel Forge — Agent harness that uses LLMs to generate and optimize CUDA kernels for ML workloads (matmul, convolution, normalization). Automates the expert-intensive kernel tuning loop. If it delivers, this could shrink the gap between prototype kernels and hand-tuned vendor libraries.

ProcAgent — Agentic framework for procedural tasks (assembly, repair) running on edge with human-in-the-loop. Handles instruction interpretation, progress tracking, spatial reasoning, error recovery — all without cloud. Targets industrial/field use cases where connectivity is spotty and mistakes are expensive.

RoCo-ACE — Rollout-conditioned online distillation for retention-aware knowledge injection in MLLMs. Fine-grained supervision on important tokens during distillation to prevent drift in non-updated behaviors. The “retention-aware” framing is the key — most knowledge injection catastrophically forgets.

Right-sizing Recommendations (RSR) — Conformal prediction for VM size selection in hyperscale clouds. Uncertainty-aware recommendations instead of point estimates. Cloud ops folks: this directly addresses the “guess the instance size” problem that burns budget.

AI Dev Tools

Hugging Face speech-to-speech — Open-source framework for local voice agents using open models. End-to-end speech-to-speech pipelines, no cloud required. Real-time voice interaction for privacy-sensitive apps. Finally, a voice stack you can run on your own iron.

Claude-Mem — Persistent cross-session memory for coding agents. Captures session activity, compresses with AI, injects relevant context into future sessions. Works across Claude Code, OpenClaw, Codex, Gemini, Hermes, Copilot, OpenCode. ChromaDB backend. The “works across everything” claim is the hook — if it actually works, it solves the context fragmentation problem.

anthropics/claude-code v2.1.219 — Claude Opus 5 now default (1M context, $10/$50 per Mtok fast mode). Sandbox network allowlist to deny non-allowlisted hosts without prompting. DirectoryAdded hook fires after /add-dir or SDK calls. The pricing on Opus 5 fast mode is notable — half the previous Opus rate.

sst/opencode v1.18.8 — Better MCP server compatibility and OAuth flows. Fixed reconnection after expired SDK sessions (including concurrent requests). Honors configured MCP OAuth callback ports. Dropped deprecated sampling defaults for newer Gemini models. Incremental but the MCP stability fixes matter if you’re running agent workflows in production.

cline/cline Desktop v0.0.7 — System tray with session count, paginated session history (10 per page, lazy load), favorited sessions, consistent status colors. Subagent/teammate runs now visible in session history. Quality-of-life updates that signal the desktop app is maturing past “side project” stage.

Today’s Synthesis

If you’re building an internal AI coding platform, OpenWork gives you the self-hosted collaborative foundation (built on opencode), Claude-Mem solves the cross-session memory fragmentation across Claude Code, OpenCode, Codex, and Gemini, and opencode v1.18.8’s MCP stability fixes mean your agent tooling won’t flake under concurrent production workloads. Wire them together: deploy OpenWork on your infra, point it at your opencode fork with the MCP reconnection fixes, add Claude-Mem’s ChromaDB backend for persistent context that survives session rotation, and you’ve got a coding environment that remembers decisions across weeks — not hours — without shipping a single token to Anthropic. The sandbox network allowlist in Claude Code v2.1.219 is a nice bonus if you’re running Opus 5 fast mode ($5/Mtok) and need egress control. This isn’t a demo stack; it’s three production-ready pieces that slot together today.