Agent Toolkit Search Plugin

Stateless Streamable HTTP MCP server for web retrieval and cited research. The application targets .NET 10, uses the official MCP C# SDK, and optionally synthesizes research with Microsoft Agent Framework.

Tools

`fetch`

Fetch one public HTTP or HTTPS URL.

Parameters:

url: absolute public URL.
format: raw, text, markdown, or readable; defaults to raw.
max_length: maximum returned characters, from 1 to 50,000; defaults to 50,000.
start_index: character offset for continuing a truncated response; defaults to 0.

markdown converts common HTML structure such as headings, paragraphs, lists, and links to Markdown. readable removes scripts, styles, tags, comments, and excess whitespace for article-like text.

`crawl`

Bounded same-origin crawl returning Markdown.

url: starting public URL.
max_pages: 1 to 20 pages; defaults to 5.
max_depth: 0 to 3 link depth; defaults to 1.

Crawl deduplicates URLs, stays on the starting origin, follows Markdown links, and checks /robots.txt before visiting pages. Missing or unavailable robots metadata allows crawling; matching User-agent: * Disallow rules block paths.

`search`

Search with tavily or bing and return Markdown containing titles, source links, snippets/raw content, and optional fetched page content.

query: non-empty search query.
provider: tavily or bing; defaults to tavily.
max_results: 1 to 10; defaults to 5.
fetch_results: fetch result pages as readable content; defaults to true.

`discover_llms_txt`

Checks and fetches /llms.txt, then /.well-known/llms.txt, returning Markdown. Requests use the same public-URL validation and GET-only fetch path.

`research`

Searches the web, enriches results, and returns cited Markdown. When OPENAI_API_KEY is configured, Microsoft Agent Framework synthesizes the source dossier; otherwise the dossier is returned deterministically.

Read-only and security behavior

Target pages, redirects, crawled pages, robots files, and llms.txt are requested with HTTP GET only.
The fetch tool exposes no arbitrary HTTP method and never sends target-site POST, PUT, PATCH, or DELETE requests.
Localhost, loopback, private IPv4, link-local, unspecified, and credential-bearing URLs are rejected.
Redirects are validated and limited to five hops.
Responses are limited to 10 MiB and HTTP clients time out after 30 seconds.
Crawl is same-origin and bounded by page/depth limits.

The MCP Streamable HTTP protocol itself may use its protocol-defined requests. Tavily's official search API requires an internal POST; this only calls Tavily's API and is not a page fetch.

Tavily integration is implemented by the shared SDK in shared/tavily/AgentSkillsMcp.Tavily.csproj, based on Tavily's OpenAPI specification. It provides typed Search and Extract operations plus Microsoft.Extensions.AI AIFunction wrappers.

Configuration

Set provider credentials through environment variables. Do not commit keys to source control.

Search__EnableTavily=false
TAVILY_API_KEY=your-tavily-key

Tavily is disabled by default. Set Search__EnableTavily=true before using its API key. The GitHub Actions workflow maps repository variable ENABLE_TAVILY_SEARCH to this option.

Local development

dotnet run --project AgentSkillsMcp.Search.csproj

The default transport is stdio. Use --http, --transport=http, or --transport http to host the HTTP endpoint:

dotnet run --project AgentSkillsMcp.Search.csproj -- --http

Search__Transport=http can also select HTTP for hosted environments.

Connect an MCP client to the HTTP endpoint:

{
  "mcpServers": {
    "agent-toolkit-search": {
      "type": "http",
      "url": "http://localhost:6243"
    }
  }
}

Docker

The dedicated Dockerfile uses an architecture-aware multi-stage .NET build, cached restore/publish layers, the ASP.NET runtime image, and the non-root $APP_UID user.

docker build -f mcps/search/Dockerfile -t agent-toolkit-mcp-fetch .
docker run --rm -p 8080:8080 \
  -e TAVILY_API_KEY \
  agent-toolkit-mcp-fetch

The MCP endpoint is available at http://localhost:8080.

Development validation

dotnet build AgentSkillsMcp.Search.csproj
dotnet test ../../../tests/search/unit-tests/AgentSkillsMcp.Search.Tests.csproj
dotnet test ../../../tests/search/integration-tests/AgentSkillsMcp.Search.IntegrationTests.csproj

Tests use xUnit v3, Microsoft Testing Platform, and Shouldly.

More information

ASP.NET Core MCP servers use the ModelContextProtocol.AspNetCore package from the MCP C# SDK. For more information about MCP: