dsh points web_search at DeepSeek's paid API by default: 402 on a free account, exit code 0 anyway. Tested 2026-09-18, plus a free Tavily backend.

Four free API keys, all returning 200 from curl, and not one working agent. The five failures that are never the key, and a six-step checklist.

Hands-on DeepSeek Harness (dsh) review: 205k stars in 18 days. We ran real agent tasks with four free models. Works well, but headless usage is undocumented and Groq fails silently.

LiveKit Agents (Apache 2.0) vs Pipecat (BSD-2): free tiers, runnable code, and which STT, LLM and TTS you can ship at $0 without a licence problem.

The official repo now maintains 7 MCP servers and archived 13, so most npx commands in listicles are dead. What's actually free, and where $0 stops.

Ragas, DeepEval and Promptfoo are all free and open source. Compare what each grades, and how to run every eval at $0 with a local or free-tier judge.

E2B gives $100 in credits, Daytona $200, both one-time. Real free-tier numbers, per-second cost math, and the self-hosted path with no meter.

Fine-tune an 8B LLM in ~6 GB of VRAM on a free Colab T4 with Unsloth. Official VRAM table, runnable QLoRA code, GGUF export and free-GPU options.

Browser Use is the free, MIT-licensed framework that lets an AI agent drive a real browser, no selectors. Run it at $0 with a local or free model.

Mem0, Zep, and Letta compared: exact free-tier limits, temporal graph vs vector memory vs self-editing blocks, runnable Python, and which to pick.