Agent Search MCP

Find the sources. Bring back the evidence.

Agent Search MCP is an open-source MCP web search router. It searches Chinese and English pages without a search API key by default. Search sources, permission to spend, and response length should be controllable. Agent Search MCP brings those choices into one entrypoint.

In this article

A query with its constraints attached.

A traveler studies a map beside two mountain paths leading to distant pavilions.

Illustrative request

Query · language · source policy · budget

Find Chinese technical sources, prefer free providers, and limit the evidence returned.

Evidence path

Sources → deduplication → traceable results

Read the result, follow its sources, and inspect provider failures and routing decisions.

Illustrative request: input constraints and response information, with no live search performed.

How it works

For agents that need web evidence

A query with clear constraints

Language, source policy, and budgets for calls, time, and evidence.

Zero-key sources

The default free_first policy searches these sources.

Optional paid sources

Participate when credentials are configured and policy permits.

Assemble traceable evidence

Merge and deduplicate results, retaining relevant content and sources within the evidence budget.

Return to the agent

Read results, follow original sources, and inspect failures and routing decisions.

The default free_first policy uses only zero-key sources. Policy determines how paid sources participate; configuring credentials alone does not authorize spending.Inspect the source policy

Token numbers come from a replayable fixture.

The repository replays the same bilingual results with a locked tokenizer and measures only the effect of output formatting.

OutputAverage tokens per queryVersus Normal
Normal2311.0Baseline
Compact1655.828.4%
Compact+1607.530.4%

This fixture verifies formatting and evidence-packet behavior. It does not measure live engine availability, search accuracy, or a fixed saving for every query.

Read the benchmark method

Installation and support

The default configuration uses only zero-key sources. Connect an MCP client first, then add a commercial provider, HTTP mode, or persistent cache when you need it.

Requires Node.js 18.17 or newer. The default runtime does not require Python.

Terminal
npx -y agent-search-mcp
MCP client configuration
{
  "mcpServers": {
    "agent-search": {
      "command": "npx",
      "args": ["-y", "agent-search-mcp"]
    }
  }
}

Install guide on GitHub

Common questions

Does Agent Search MCP require an API key?

No. Eight sources start without a key. Brave, Tavily, Exa, and You.com run only after you provide credentials and select a routing policy that permits them.

Can it replace Tavily or Exa?

It fits free-first, self-hosted, and multi-source evidence workflows. Commercial search APIs still have advantages in stability, proprietary indexes, or semantic retrieval, so the project keeps them as controlled escalation paths.

Does the default runtime need Python?

No. Default search runs on Node.js. Only the optional semantic deduplication and reranking features use Python with Model2Vec.

How does it reduce agent context use?

Compact output keeps full detail for the first results, turns later results into source-preserving references, and caps query-relevant passages with one evidence budget.

Is HTTP mode secure by default?

HTTP and both modes require a Bearer token by default. Browser requests with an Origin header must also match the allowlist. stdio remains the simplest local path.

Install guide on GitHub

Keep reading

Selected by You.com: How a Bot PR Became an Architecture Validation

Free Tavily Alternative: Agent Search MCP Web Search

AI Agent Search API Buyer's Guide: Brave vs Tavily vs Exa vs DDG

View the source

Back to the project story