AI For Dev

summary-bart-slodyczka-gemma4-searxng-openclaw

3 min read

Gemma 4 + SearXNG = 100% FREE & PRIVATE OpenClaw (Full Setup)

Channel: Bart Slodyczka Format: YouTube video URL: https://www.youtube.com/watch?v=T0CKsU0hQx4 Length: ~12 minutes Published: 2026-04-06 Sponsor: none disclosed

Summary

Step-by-step guide for running OpenClaw 100% locally and free using Gemma 4 models via Ollama, plus SearXNG as a self-hosted web search provider via Docker. The combination delivers a fully private personal AI agent with no rate limits and no data leaving the device.

Key Points

  • E4B tool-calling impressed Bart — he ran the E4B model (mobile-grade, ~9.6 GB on disk) as his OpenClaw daily driver for several days and found it “very impressed with its ability to do agentic tool calling” — previously small models would “buckle halfway through” multi-step tasks. OpenClaw’s improved backend prompt structure gets more out of smaller models now.
  • 26B is Bart’s current daily driver — running the 18 GB version on a 512 GB Mac Studio; also fits comfortably on a 24 GB MacBook Pro. The 256K context window is the sweet spot for sustained OpenClaw conversations.
  • E2B tested for 16 GB machines — 7.2 GB on disk; viable for 16 GB MacBook Pro users, though the remaining headroom for conversation context is tight.
  • Ollama native integration with OpenClaw — openclaw configure → select Ollama → models auto-detected from ollama list. No custom coding. After changing settings, openclaw gateway restart to flush.
  • Ollama Cloud offering — $20/month plan for users on VPS or small devices. Runs large models (Kimi K2.5, GLM 5, MiniMax M2.7) in the cloud. Can also run the Gemma 4 31B with full 256K context via cloud instance.
  • SearXNG — self-hosted meta-search engine running in Docker with native OpenClaw integration. Setup: docker pull + one config change (add json format to settings.yaml under the formats section — OpenClaw uses JSON, SearXNG defaults to HTML only). Fully private: no queries leave the machine.
  • First-message latency — 10–20 seconds for first response after a cold start (loading model into RAM); subsequent responses faster in a warm conversation.
  • Claude Code as setup assistant — Bart suggests spinning up a Claude Code terminal instance to read your OpenClaw + Docker configs and auto-configure SearXNG, rather than doing the YAML editing manually.

Sponsorship & Bias Notes

Sponsor: none disclosed. Product placement / affiliations: Bart mentions Ollama Cloud ($20/mo) as an option but doesn’t appear to have an affiliate relationship. None observed. Comparison bias: none observed.

Connected Pages

See Also

Graph View

Backlinks