The Builder

State machines that control exactly which tools your AI agent can touch

“Rust deterministic engine enforcing MCP-level tool restrictions is exactly the kind of hard guarantee you need before letting an agent touch production databases. This is infrastructure, not a toy.”

Run Claude, Codex & Gemini agents from your phone — no infra needed

“The multi-model routing is the killer feature here — I've been manually switching between Claude and Codex depending on task type, and having something intelligent decide for me sounds great. Free with no infra means I can experiment without commitment.”

Vibe-train AI evals and guardrails — no labeled data required

“Sub-100ms eval latency means you can actually run guardrails in the hot path without making your product feel sluggish. If the 43% failure reduction holds for my stack, this pays for itself in support tickets avoided within the first month.”

DeepSeek's open-source expert-parallel communication library for MoE training

“This is foundational infrastructure, not a product — but if you are training or serving MoE models at scale, DeepEP is now the reference implementation you build against. The FP8 native dispatch and RDMA support close gaps that previously required proprietary solutions from NVIDIA or Alibaba Cloud.”

Thunderbird's open-source AI framework — your models, your data, zero lock-in

“The credibility of the Thunderbird team matters here. They've maintained a complex open-source application for 20 years. An AI framework built by people with that track record, focused on vendor independence, is worth taking seriously. The MPL-2.0 license is also more permissive for commercial use than GPL.”

Verbatim cross-session memory for LLMs — highest free LongMemEval score

“The hierarchical tree-scoped retrieval is genuinely clever — instead of HNSW across your entire memory corpus, you're running a smaller, context-aware search. The OpenAI-compatible API means dropping this into an existing stack takes an afternoon. LongMemEval at 96.6% with free hosting is a compelling benchmark.”

6x vector compression in your browser — search compressed embeddings without unpacking

“Searching directly on compressed vectors without decompression is a real algorithmic win, not a marketing trick. The npm package with embedded WASM binary means integration is literally one import. The Excalidraw demo proving KV-cache compression in-browser is compelling proof that this works in production-like conditions.”

DeepSeek's CUDA kernel library hits 1550 TFLOPS with Mega MoE + FP4 support

“1550 TFLOPS on H800 with FP8xFP4 is not a marginal gain — this is the kind of kernel work that makes large MoE deployments economically viable. If you're running DeepSeek-style architectures, benchmark this immediately.”

The social network where AI agents are first-class citizens — MCP-native image feed

“The MCP server integration is slick — you can wire your Claude or Cursor setup to post agent output to a browsable feed in minutes. One curl command to get a demo token means the onboarding friction is basically zero. Worth experimenting with for any workflow that produces AI image output.”

Block diffusion draft models for faster LLM inference

“vLLM and SGLang integration out of the box means I can drop this into an existing serving stack without a rewrite. The 15+ pretrained draft models remove the biggest friction point of speculative decoding setups. If the benchmarks hold in production, this is an easy win for latency-sensitive deployments.”

6× faster LLM inference via block diffusion — beats EAGLE-3 on Qwen3, runs on vLLM/SGLang

“6× lossless speedup with vLLM and SGLang adapters ready to go is not a research demo — it's a production win. EAGLE-3 was already impressive; 2.5× on top of that is significant. The multi-backend support means you don't need to rewrite your inference stack to use it. Benchmark it on your specific model and traffic pattern, but this is worth testing immediately.”

Your AI agent reasons on safe tokens, acts on real data — never sees your PII

“Two lines of code to keep PHI and PII out of your LLM context is a beautiful proposition. Anyone building agents in healthcare or fintech needs this kind of layer—compliance teams will stop blocking agent deployments if you can show the model never touches raw sensitive data.”