Vai al contenuto principale
T

argo (dsh-plugin)

taxueseek/argo/packages/dsh-plugin

Ricerca unificata con verifica delle evidenze per agenti: oltre 120 motori con instradamento multilingue, domini verticali e schede di modalità strutturate, 10 strumenti MCP più orchestrazione wide_research e un ciclo fetch-verify che completa i punteggi delle evidenze.

Installazione

dsh plugin --profile web add github:taxueseek/argo

Questo plugin si trova nella sottodirectory packages/dsh-plugin del repository.

README

Argo: unified search and evidence verification for AI agents across hundreds of sources and 18 languages

English · 中文 · 日本語 · 한국어 · Español

Quick Start · Why Argo · Routing Proof · How It Works · AI Search & Sub-skills · DSH Plugin · Docs

stars npm downloads skills.sh license python version engines mcp

Part of the taxueseek skill series (search & deep research): plainspoken-writing 75★ (AI-tone removal) · session-digger 34★ (cross-environment session mining) · taxue-invest 30★ (investment analysis) · taxueskills 25★ (decision system) · taxue-creative-style 19★ (creative illustration styles) · ime-lexicon 15★ (IME lexicon toolkit).

DeepSeek Harness (DSH) plugin ecosystem: dsh-files 43★ (folder upload + document parsing) · plugin-fast-adapt 15★ (plugin upgrade adapter) · taxue-dsh-artisan 3★ (visual creation) · dsh-plugin-guard 2★ (pre-install audit).

Argo is an open-source metasearch, web-fetch, and evidence-verification engine that lives inside your AI agent. It routes queries across hundreds of sources in 18 languages (web, academic, finance, code, local files, and leading AI search engines), falls back to keyless & local engines when API keys are absent, and verifies facts from full page text instead of trusting search snippets.


Quick Start

Zero-config works out of the box (API keys are optional). Pick your client:

  • Claude Code (MCP):
    claude mcp add argo -- npx -y argo-search
    
  • Agent Skill (skills.sh — Claude Code, Cursor, Codex, Antigravity):
    npx skills add taxueseek/argo
    # Or install all scenario skills (argo, argo-deep-research, argo-local-search):
    npx skills add taxueseek/argo --full-depth --all
    
  • Cursor / Windsurf / LibreChat (mcpServers JSON):
    {
      "mcpServers": {
        "argo": { "command": "npx", "args": ["-y", "argo-search"] }
      }
    }
    
  • Local CLI (recommended for long-term local use):
    curl -fsSL https://raw.githubusercontent.com/taxueseek/argo/main/scripts/install.sh | bash
    # Windows (PowerShell): irm https://raw.githubusercontent.com/taxueseek/argo/main/scripts/install.ps1 | iex
    # or via npm: npm install -g argo-search
    
  • DeepSeek Harness (DSH) Plugin:
    dsh plugin --profile web add argo-search
    
Verify installation, requirements & optional API keys
argo search "python asyncio error handling" --verify 2 --json --fields agent
argo search --list-engines
  • Requires Python 3.9+ (pip install pyyaml) and Node.js 18+ (for npx).
  • Optional keys (DEEPSEEK_API_KEY, ARGO_YUANBAO_API_KEY, ARGO_TAVILY_API_KEY, EXA_API_KEY, ARGO_METASO_API_KEY, ARGO_BOCHA_API_KEY) improve quality when present and are skipped automatically when missing. See docs/MCP_SETUP.md and docs/ENGINE_CATALOG.md.

Why Argo

Why Argo outperforms built-in search, single AI search APIs, and generic metasearch

CapabilityBuilt-in Web SearchCommercial Search APIs (Tavily / Exa / Serper)Argo
Pricing & KeysTied to host quotaPaid per call; API key requiredFree & keyless fallback (optional keys boost quality)
AI Search AggregationSingle host backendSingle vendor indexUnifies keyless (tavily_free, parallel_free, anysearch) + BYOK (DeepSeek Search, Tencent Yuanbao/Hunyuan, Metaso, Bocha, Tavily, Exa)
Source RoutingSingle generic indexGeneric web indexDomain + 18-language routing across academic, finance, code, news & local files
Fact VerificationSnippet onlySnippet or raw chunksBuilt-in --verify & --numverify (fetches source & scores credibility)
Context EfficiencyVerbose HTML/JSONRaw chunksAgent-compact output (--fields agent ~4.3 KB)

Query-Shaped Routing

Argo query-shaped routing across finance, academic, code, and general domains

Argo classifies each query by domain and language before dispatching, hitting first-party sources directly and triggering early-stop once consensus is reached:

  • Academic → arxiv, semantic_scholar, openalex, crossref, pubmed (--academic-deep)
  • Finance & Macro → sec_edgar, yahoo_finance, sina_quote, eastmoney
  • Code & Tech → stackoverflow, github, octen, context7
  • AI Search → tavily_free, parallel_free, anysearch, deepseek_search, yuanbao, metaso, bocha, exa, tavily, you
  • Deep Research → python3 scripts/research.py "topic" --depth deep --json outputs a structured dossier (sources, coverage gaps, quality gates)

How It Works

Argo retrieval, RRF fusion, fetch degradation chain, and evidence scoring pipeline

Evidence Scoring (Selection × Absorption)

Search engines optimize for clicks; agents need verifiable facts. Argo scores every candidate along two orthogonal axes:

  • Selection: domain authority, consensus across engines (RRF), freshness, and SERP-shell rejection (blocks baidu/s and sogou/link redirect stubs from posing as primary sources).
  • Absorption: actual readable density after fetching the full page (--verify N) and numerical cross-checking (--numverify N for prices, financials, and dates).

Budget Modes (--mode)

ModeBehaviorBest for
fastFree + local local_* engines only, zero API spendHigh-frequency agent lookups
autoCost-aware default: free first, escalates to keyed sources when neededDaily coding & general research
deepBroad multi-source fan-out, disables early-stop, pairs with --verifyHigh-stakes finance, legal, academic dossiers
budgetStrict daily/monthly quota caps with local fallbackUnattended batch & cron workflows

AI Search & Sub-skills

Sub-skills (7, Single Directory)

All sub-skills live in sub-skills/. Four are built-in pipelines driven by the main skill; three are standalone entry points (wide research orchestration, and two scenario bundles for deep research / zero-cost local search).

Sub-skillCapability
local-searchLocal file search + 32 zero-cost local aggregation engines (--local-first)
local-seekHybrid BM25 + vector semantic search for local notes, markdown, and knowledge bases
ego-searchReal-browser logged-in search for Zhihu, Xiaohongshu, WeChat Official Accounts (off by default)
js-runSub-millisecond headless JS execution for environment probes and pure-compute pages
argo-wide-researchMulti-track parallel deep research & forensics orchestration with citation ledger and quality gates
argo-deep-researchScenario entry: research dossier + multi-source fact/number verification (research.py, evidence.py, --verify)
argo-local-searchScenario entry: zero-cost keyless search bundle (--local-first, seek.py, --include-local)

19 MCP Tools

argo_search, argo_local_search, argo_local_read, argo_recompute, argo_research, argo_evidence, argo_clarify, argo_fetch, argo_crawl, argo_screenshot, argo_pdf, argo_social_search, argo_article, argo_job, argo_extract, argo_preflight, argo_answer, argo_watch, argo_cite.


DSH Plugin

Argo ships a first-class DeepSeek Harness (DSH) plugin (@taxueseek/argo-dsh) with three invocation channels from one capability set:

  1. 19 mcp__argo__* tools (full MCP bundle; default MCP surface exposes 4 core tools, set ARGO_MCP_TOOLS=all for all 19).
  2. Search + wide_research parallel orchestration (subpackage).
  3. Host web_search / web_fetch seam replacement (requires explicit provider pin in the web profile; see packages/dsh-plugin/README.md).

The built-in settings UI exposes 18 native first-class tools, 35 behaviour switches, and shares DSH's native DEEPSEEK_API_KEY credential store with zero plaintext echoing:

DSH Tools settings tab DSH Engines settings tab DSH Keys settings tab


Docs & Changelog

VersionHighlights
v2.15.0Zhihu full chain, evidence-layer authority passthrough, keyless 3-tier semantic fallback, MCP generic connector (type: mcp), route zero-based domain scoring, 311 sources
v2.14.1Architecture slimming, MCP 3+1 orthogonal tools + backward-compat aliases, 304 sources, fail-closed secret_guard.py redaction
v2.14.0Parser layer consolidation, 300 sources, badcase ledger loop
v2.13.0Documentation consolidation, 22-state-machine SSOT, dispatch tail grace
v2.15.0 Release Summary & Capability Snapshot

Capabilities (v2.15.0)

v2.15.0 delivers the Zhihu full chain (question drill-down, curated answers, Keychain credentials, vote-quality weighting), evidence-layer upstream_authority passthrough, a keyless 3-tier semantic fallback (local 2-gram heuristics), the single-file research-report pipeline, route zero-based domain scoring, and the MCP generic connector (type: mcp, first integration grep.app). Sources 304→311, domains 306→313.

v2.15.0: Zhihu Full Chain + Evidence Authority Passthrough + MCP Connector

See docs/RELEASE_NOTES_v2.15.0.md and docs/CHANGELOG.md.

v2.14.1: Architecture Slimming + MCP 3+1 Tools + Security Redaction

See docs/RELEASE_NOTES_v2.14.1.md and docs/CHANGELOG.md.


Good search is not about seeing more—it is about concluding with confidence, and knowing when you still should not.

Star History Chart

If Argo has ever helped you reach a more confident conclusion, give it a Star to help more people find it.

Contributing & License

Issues and PRs are welcome (see docs/ADDING_NEW_ENGINE.md to add a source). Licensed under Apache-2.0.

Plugin correlati