Argos Engine
Token-efficient, local-first web research MCP server in Rust. Keyless multi-engine search with parallel fanout, dedup and automatic failover — no API keys, no runtime, no VM required. Built on FlojoMCP over the official Rust MCP SDK (rmcp).
Keywords: MCP server, web search, deep research, metasearch, keyless, token efficiency, AI agents, Model Context Protocol, local-first, Rust, DuckDuckGo, Bing.
Why
Search MCPs force a bad trade: scrape fragile engines yourself, rent an API key, or run a container farm — and either way the agent pays for it in tokens and timeouts.
- Keyless by default: DuckDuckGo, Bing and Brave HTML endpoints queried directly from Rust for compatibility. No accounts, no cards, no Docker, no WSL. The direct HTML set is experimental: see the measured quality and policy caveats below before treating it as a durable index.
- Browser impersonation: the HTTP layer (primp, same crate ddgs uses) masquerades as Chrome153/Windows — measured difference: without it Bing serves a
302to its homepage and Brave returns429; with it all three engines answer200. - Fanout with failover: all providers queried in parallel, results fused with weighted RRF and deduplicated by canonical URL. When one engine rate-limits your IP (202/403/429 happen — see below), the others still answer: research degrades, it doesn't die.
- Tokens are a feature: compact results, server-side truncation budgets (
src/limits.rs), dedup before the agent ever sees duplicates. - No silent failures: every
searchresponse carries the per-provider status (ok,empty,filtered,rate_limited,unreachable) and the warnings. If all engines were throttled, the tool returns a typedAllProvidersRateLimitederror; if providers answered but every result failed Argos' domain/relevance gate, it returnsNoUsableResultsinstead of pretending the topic had no hits. - Honest, typed errors:
rate_limited,unreachable,all_providers_rate_limited,no_usable_results— each with an actionable hint, never a genericError executing toolghost. - Per-provider visibility: every
searchresponse carries{ results, providers: [{name, kind}], warnings }so the agent knows exactly which engines contributed, which were empty or filtered, and which errored. - One small binary: ~7 MB exe, ~5 ms startup, ~4 MB RAM — no Python/Node runtime.
Status
Early development — Milestone 1: keyless multi-provider search. Not production-ready yet.
| More keyless providers | 📋 provider audit first; see PROVIDER_AUDIT.md |
| research tool: multi-query fan-out, progress + cancellation, compact digest | 📋 M2 |
| Content extraction (readability → markdown, no XPath) | 📋 M3 |
| Optional cloud providers (Serper/Tavily/Exa) behind SearchProvider | 📋 M3+ |
Requirements
- Rust 1.89+ (rustup) — primp's floor
- Windows: VS Build Tools (the
.batscripts set upVsDevCmd); plaincargo buildworks on any platform. - Nothing else: no Docker, no WSL, no VM, no API keys.
Quick start
cargo build --release # or build.bat on Windows (release + tests)
Connect from opencode.jsonc:
"argos": {
"type": "local",
"command": ["D:\\Mis Juegos\\ClaudeMCPs\\argos-engine\\target\\release\\argos-engine.exe"]
}
Or any MCP client (Claude Desktop, Cursor, Inspector…): command = the argos-engine binary, stdio transport.
Tools
Both tools declare all four MCP annotations, so a client can tell before invoking that they are read-only, non-destructive, idempotent and open-world:
{ "readOnlyHint": true, "destructiveHint": false, "idempotentHint": true, "openWorldHint": true }
search only reads: nothing the caller can observe is mutated, so a host can invoke it without warning.
Configuration
The rate-limit reality (measured, not guessed)
Anti-bot limits live in the upstream engines, not in your client — and every keyless approach shares that wall — SearXNG's own docs admit it gets CAPTCHAs too, which is why the optional adapter was dropped. Measurements from one residential IP,2026-09-24:
Argos raises the ceiling with design instead of hoping: browser impersonation at the transport layer, parallel fanout across independent rate budgets, weighted RRF fusion + canonical URL dedup (fewer redundant queries for the agent), a short-lived cache with single-flight, per-request timeouts and typed rate-limit errors with failover.
Quality reality (measured 2026-09-24)
HTTP 200 is not a quality result. A controlled live audit found that DDG and Brave can return highly relevant niche results on their first request and then quickly return 202/429 from the same IP. Bing returned valid HTML with b_algo entries but poor semantic coverage for the same queries: free piano VST produced free-game pages, and one Spitfire Audio LABS free run produced Macao travel pages. The query itself was present in Bing's final URL, so this was not a URL-encoding failure.
Argos therefore applies a conservative result-side quality gate:
domainsis validated and enforced against the returned host;site:text is only a ranking hint, never the security boundary.- Results with no meaningful query terms are rejected as obvious entity drift.
- A provider whose entire result set was rejected is
filtered, notok. - If no provider contributes a usable result, the tool returns typed
NoUsableResults.
The full matrix and policy notes are in PROVIDER_AUDIT.md. The direct HTML endpoints remain a compatibility set, not a guarantee of quality or long-term provider availability.
Comparison (researched 2026-09-23)
¹ FastMCP/Python figures from FlojoMCP benchmarks; ddgs behavior from source inspection of v9.16.0.
Roadmap
- M1 — keyless fanout (DDG + Bing + Brave, primp impersonation), dedup, typed errors, fixtures, per-provider outcome and validated
domainsscoping ← current - M1.1 — provider quality gates, result-side scope enforcement and
NoUsableResults; see PROVIDER_AUDIT.md ← current - M1.2 — Native Metasearch Engine: provider registry, profiles, index-family diversity, circuit breaker, quota/cooldown state, RRF fusion and compact token-aware output; see PROVIDERS.md ← current
- M1.3 — keyless and public API adapters (academic, code, news and local verticals)
- M1.4 — user-key/free-tier adapters, with provider-owned keys and quotas
- M2 —
researchtool: multi-query fan-out with reformulation, optional source allowlist, progress + cancellation, compact digest (ChatGPT/Qwen deep-research style) - M3 — content extraction (Rust readability, no XPath) and optional structured remote providers behind the same trait
Decisions and progress live in AGENTS.md.
License
GPL-3.0 — see LICENSE.