laxmi-search-router

Multi-provider web search HTTP service for the LAxMI fund. Forked from robbyczgw-cla/hermes-web-search-plus — Hermes plugin shim removed, replaced with a fund-style HTTP service so any LAxMI bot (jekyl/hyde/scout/herald/etc.) can call it directly.

What changed vs. upstream

	Upstream	This fork
Entry point	Hermes plugin (__init__.py:register)	HTTP server (server.py on :3451)
Env loading	reads plugin-local .env	inherits from dotenvx-decrypted fund env
Halt awareness	none	reads /tmp/fund-state.json, refuses while halted
Caller telemetry	none	exposes search/extract counts via GET /
Routing engine	search.py	unchanged — all credit to upstream for the auto-router

Running

# Standalone
~/.venvs/search-router/bin/python /Users/shaway/laxmi-search-router/server.py

# Via fund.mjs (Phase 1 opt-in — see fund.mjs BOTS array)
node fund.mjs

Provider API keys (SERPER_API_KEY, BRAVE_API_KEY, TAVILY_API_KEY, EXA_API_KEY, QUERIT_API_KEY, LINKUP_API_KEY, FIRECRAWL_API_KEY, PERPLEXITY_API_KEY, YOU_API_KEY, SEARXNG_INSTANCE_URL) are read from process env. In fund context they come from the encrypted .env via dotenvx. At least one is required; missing keys just disable that provider.

API

GET / — status

{
  "bot": "search-router",
  "cycle": 14,
  "running": false,
  "last": {"kind": "search", "query": "...", "provider": "tavily", "result_count": 5, "cached": false},
  "providers_configured": ["serper", "brave", "tavily", "exa"],
  "counts": {"search": 12, "extract": 2}
}

POST /search

curl -s localhost:3451/search -H 'Content-Type: application/json' -d '{
  "query": "Aerodrome USDC pool TVL",
  "provider": "auto",
  "count": 5,
  "time_range": "week"
}' | jq

Body fields: query (required), provider (auto/serper/brave/tavily/exa/querit/linkup/firecrawl/perplexity/you/searxng), count (1–20), depth (normal/deep/deep-reasoning — Exa only), time_range (day/week/month/year), include_domains[], exclude_domains[].

Returns the underlying search.py JSON: {provider, results: [...], routing: {...}, cached, ...}.

POST /extract

curl -s localhost:3451/extract -H 'Content-Type: application/json' -d '{
  "urls": ["https://example.com/article"],
  "provider": "firecrawl",
  "format": "markdown"
}' | jq

POST /trigger

No-op (request/response service has no cycle). Returns {ok: true} so the fund status aggregator and ops-chat treat it like any other bot.

Calling from a fund bot

const res = await fetch("http://localhost:3451/search", {
  method: "POST",
  headers: { "Content-Type": "application/json" },
  body: JSON.stringify({ query, count: 5, time_range: "week" }),
});
const { results, routing } = await res.json();

Python equivalent (see limitless-bot.py style):

import requests
r = requests.post("http://localhost:3451/search",
                  json={"query": q, "count": 5}, timeout=80)
data = r.json()

Fund integration notes

• Port: 3451 (next free slot in CLAUDE.md ports table — bump that table to 3452 when this lands).
• Halt-aware: while /tmp/fund-state.json.halted == true, /search and /extract return 503 with the halt reason. Status endpoint stays up.
• No money moved: read-only API calls only. No DRY_RUN gate needed.
• Cost: free tiers cover normal usage. Heavy callers (deep research, mass extract) can blow through Tavily/Exa monthly limits — provider cooldowns kick in at 1h on failure.
• Phase: opt-in. Phase 0 doesn't need it; reactivated bots in Phase 1+ (jekyl/hyde/scout) benefit most.

Credit

Routing engine (search.py, ~3.4k lines) is unchanged from upstream. All provider integrations, scoring heuristics, cooldowns, and caching logic are the original author's work. This fork only adds the fund's HTTP/halt/telemetry shell around it.

License

MIT — same as upstream.