As of October 26, 2026, the 5 AI tools below were all newly launched or received major updates within the past week, spanning voice agents, model-layer memory, low-cost GPU inference, enterprise LLM gateways, and AI social-video studios — every one has a verifiable project page.

1. Lampline

Overview: An open-source toolkit (Apache-2.0, nearly 300 stars in its first week) for building AI phone agents on your own computer. The killer feature is “using AI to test AI”: after every change, an AI caller dials your agent with a real phone call to validate the flow — no real phone number and no API keys needed to run the dev loop, and once you plug in your own voice provider, each test call costs cents. It supports SIP, speech-to-text/TTS, IVR, and importing Dialogflow CX flows, and is fully self-hostable — a fast iteration loop for customer-service bots, outbound campaigns, and voice ordering.

Pricing: Open source, free (Apache-2.0); third-party voice/LLM services billed separately.

Link: https://github.com/C4rohan/Lampline

2. MemT 1.0 (MemTensor)

Overview: One of the hottest model-layer memory open-source projects on GitHub this week (~400 stars). Unlike the usual “vector database + external retrieval” stack, MemT 1.0 builds memory directly into the model layer — giving LLMs native long-term memory rather than stitching context together per session. The official repo supports mainstream open models like DeepSeek and Qwen, with adaptation for the sglang inference framework; QbitAI covered its technical route last week under the headline “Adding the ‘memory’ piece at the model layer.” For developers building agents, multi-turn conversation, and long-context applications, it’s one of the few open attempts at “memory in the weights.”

Pricing: Open source, free.

Link: https://github.com/MemTensor/MemT-1.0

3. kindi

Overview: A custom LLM inference engine (AGPL-3.0) built for aging, low-cost GPUs. The author’s goal: let Quadro K420 and K1200 (even two K1200s joined into one 8 GB card), a GTX 1080, or an RTX 3080 run modern small models — via ternary/bitnet quantization and tensor parallelism, serving Bonsai 2B/4B/9B/27B models with an OpenAI-compatible endpoint. If you have a dusty old GPU lying around and want a usable local LLM, this is one of the most aggressively cost-optimized routes out there.

Pricing: Open source, free (AGPL-3.0).

Link: https://github.com/ahtx/kindi

4. xhub (Qiniu)

Overview: A self-hosted enterprise AI gateway from Qiniu Cloud (MIT License, first week live). One OpenAI-compatible endpoint unifies 100+ LLM APIs (Bedrock, Azure, OpenAI, Anthropic, VertexAI, vLLM, Nvidia NIM, and more) with built-in cost tracking, guardrails, load balancing, and full logging — letting enterprises consolidate scattered model calls behind one controllable entry and charge Token costs back to projects/teams. Unlike free-routing hobbyist projects, this one targets production environments that need audit, governance, and billing.

Pricing: Open source, free (MIT); self-hosted.

Link: https://github.com/qiniu/xhub

5. 37studio

Overview: An open-source AI social-media video studio (MIT), positioned as a self-hosted alternative to Buffer + Opus Clip: Claude Code produces on-brand videos (including HyperFrames workflows), publishes them to 9 platforms, and drafts reply comments automatically. Fully self-hosted with your own API keys — data never leaves your infrastructure. Go backend + Docker deployment, ships with an MCP server — a fit for content teams that want the whole “produce → distribute → engage” loop inside their own servers.

Pricing: Open source, free (MIT); usage billed via your own API keys.

Link: https://github.com/Shaitrish/37studio

All tools are subject to their latest releases; click through for live demos and docs.