Summary

Extraction that knows when it got half the article.

Completeness assessment runs after every successful extraction, scoring signals like trailing ellipses, feed truncation markers and mid-sentence endings, and returning whether the text is complete with a confidence value. Above 85% confidence it keeps trying the next extractor instead of handing back a partial page — which is the difference between a research pack built on articles and one built on RSS teasers. The other structural decision worth understanding: budgets, sessions and evidence live in the HTTP authority, so restarting the MCP adapter cannot fork the accounting.

What it is

A retrieval platform for AI agents. It routes queries across search providers by tier — free ones first, budget-exhausted ones skipped automatically — extracts full page text rather than links, recovers dead URLs, and writes what it finds into a local corpus with traceable artifacts. The MCP server is a stateless adapter over its authenticated HTTP API.

What you get

Web search with provider fusion, restrictable to free tier-0 providers only, and optional per-provider score attribution on each result — `search_web`
Full text from a URL with a completeness assessment attached, and link expansion — `extract_content`, `expand_links`
Recovery when a URL is dead, paywalled or radically changed, using Wayback Machine and archive fallbacks — `recover_url`
Three workflows that produce local artifacts, not just responses: recover a dead article with a citation-backed report, capture the important pages of a site on-domain, and build a docs-plus-research pack — `recover_dead_article`, `capture_site`, `build_research_pack`
Operational reads: provider health and remaining budgets — `search_health`, `search_budgets`
Four search modes selecting provider order within each tier: discovery, research, recovery and grounding
Multi-turn sessions through a session id, so follow-up searches are refined by earlier context

Requirements

Zero-key to start: `pip install argus-search` gives you DuckDuckGo and Yahoo with no accounts. Python 3.11 or higher. For MCP, `pipx install argus-search[mcp]` then `argus mcp init --global --client all`, which writes native configs for Claude Code, Codex CLI, OpenCode and Cursor. Production MCP delegates to an authenticated HTTP authority — set `ARGUS_AUTHORITY_URL` and `ARGUS_AUTHORITY_TOKEN`; a local broker requires `ARGUS_MCP_STANDALONE` set to true and is rejected in production. Provider keys are all optional and named per provider, including `ARGUS_BRAVE_API_KEY`, `ARGUS_TAVILY_API_KEY` and `ARGUS_WOLFRAM_API_KEY`. Remote transports need `ARGUS_API_KEY`. The package is `argus-search` (1.6.4 in pyproject); status is beta.

Setup effort

One command plus a key — pipx install 'argus-search[mcp]', then supply credentials

Argus