Spokes.wiki Search About
Software Source Code source ↗ source url updated Wed Jul 08 2026 00:00:00 GMT+0000 (Coordinated Universal Time)

ask-search

A skill (ythx-101, MIT) that gives an agent self-hosted web search with no API key. It wraps SearxNG — a privacy-respecting meta-search engine aggregating Google, Bing, DuckDuckGo, Brave, and 70+ other sources — behind an MCP server, a CLI, and an OpenClaw SKILL.md descriptor, so a coding agent gets ranked web results (with snippets and source attribution, JSON mode for programmatic use) without paying a commercial search API.

How it works

  • Self-hosted SearxNG via Docker/Docker Compose is the default backend — zero API keys, results aggregated and re-ranked across many engines, filterable by category / language / specific engine.
  • Three surfaces over one core (scripts/core.py): a CLI, an mcp/server.py MCP server for agent frameworks, and a SKILL.md for OpenClaw. Targets Claude Code, OpenClaw, and Antigravity by name.
  • Tavily API as an optional fallback when the self-hosted route isn’t wanted, plus documented workarounds for anti-bot blocking on deep-dive queries. Python 3.8+, install.sh automation.

Why it matters here

  • Extends the local-first-tool-over-MCP corner from retrieval to the live web. The spoke already tracks local-first stores reaching the agent as MCP tools — recall (memory), knowledge-rag (local document RAG). ask-search adds the live web as the same shape: a self-hosted index the agent queries as a tool, offline-of-the-vendor by default. Where knowledge-rag grounds the agent in your files, ask-search grounds it in the current web — both without a cloud API.
  • The zero-cost counterpart to a commercial search tool. from-local-llm-to-agent wired an agent to Tavily (a paid search API) over MCP as its grounding tool; ask-search is the self-hosted, no-key version of exactly that role — and keeps Tavily only as a fallback. It’s the “skills × MCP = agency” formula with the tool side made free and private.
  • Ships a SKILL.md for a specific host. Like the rest of the skill corpus it advertises itself via a SKILL.md, here scoped to OpenClaw — another data point on the skill format spreading per-host (the standardization-reach question).

Tier

T1 — first-party MIT repo, primary on itself. Young; no independent evaluation of result quality or the anti-bot workarounds (weakness recorded). freshness: volatile (depends on SearxNG + the engines it scrapes).

knowledge-rag · from-local-llm-to-agent · recall · model-context-protocol · agentskills-spec · openclaw · watch-skill · synthesis

Related pages

Links to

Linked from