Tenkai Daily — July 19, 2026

Model Releases

• prism-ml/Ternary-Bonsai-27B-mlx-2bit — A 2-bit ternary MLX quantization of Qwen3.6-27B built for on-device inference on CUDA and Metal. If you’ve got Apple Silicon and want a 27B that doesn’t eat your RAM, this is worth a look 🤖
• unsloth/inkling-GGUF — GGUF quants of Thinking Machines’ Inkling multimodal MoE that handles image-text and audio-text tasks. Usable via endpoints and local runners if you want a small multimodal model without rolling your own quant 📄

Open Source Releases

• elder-plinius/G0DM0D3 — A “liberated” AI chat client with no apparent guardrails or hosted dependencies. Read the source before you run it; the naming is doing a lot of work.
• Show HN: Recall: Give Claude memory with Redis-backed persistent context — An npm package that wires Redis-backed persistent context into Claude sessions so it stops forgetting who you are between calls 🛠️ (171 points, 93 comments)
• Show HN: Cloud-Ready Postgres MCP Server — An MCP server exposing Postgres operations, pitched as cloud-deployable for agent workflows (167 points, 79 comments)
• Show HN: MCP server for Blender that builds 3D scenes via natural language — Natural-language scene generation in Blender through MCP — useful if your agent needs to stop writing Python and start making cubes (151 points, 61 comments)
• Show HN: WhatsApp MCP Server — Lets an LLM send and read WhatsApp messages through MCP. Highest engagement of the batch, which tells you something about where people want their agents poking (229 points, 138 comments)
• Show HN: An MCP server that gives LLMs temporal awareness and time calculation — An MCP server for time math and “temporal awareness,” because apparently your model still thinks February has 30 days (91 points, 55 comments)

AI Dev Tools

• KnockOutEZ/wigolo — A local-first web research tool for coding agents over MCP — search, fetch, crawl with no API keys and no per-query cost. Public beta, so expect rough edges 🔥
• affaan-m/ECC — An “agent harness performance optimization system” with skills, memory, and security tooling for Claude Code, Codex, Opencode, and Cursor. The star count is doing impossible numbers, but the scope is broad if you’re tuning agent setups 🛠️
• shareAI-lab/learn-claude-code — A from-scratch educational “nano Claude Code” agent harness in Python, for people who’d rather read the internals than trust the black box (71k stars, apparently)
• Show HN: Mcp-Agent – Build effective agents with Model Context Protocol — A framework for composing agents around MCP servers, from LastMile AI (80 points, 28 comments)

Today’s Synthesis

The MCP gold rush is getting absurd: between the WhatsApp MCP Server , the Cloud-Ready Postgres MCP Server , and the MCP server for Blender , your agent can now text your mom, query prod, and render a donut without you writing a line of glue code. The concrete move: stop hand-rolling integrations and stand up a local MCP router (pair it with mcp-agent to compose them) so your coding agent gets Postgres + research via wigolo in one session. Just don’t wire the WhatsApp one to the same agent that has write access to your database — 229 points of internet enthusiasm is not a security review. If you’re running models locally anyway, throw Ternary-Bonsai-27B on the Metal backend and keep the whole stack on-device.