Search Rotation README

One MCP server for web search and page extraction across multiple providers.

Built by Robin Bially.

Rotate across available quotas, automatically fail over when a provider is unavailable, and return consistent results to your AI assistant. A local dashboard lets you manage API keys, reorder engines, check quotas, and inspect request history.

Actual dashboard UI, shown in English with illustrative demo data. Calls and provider budgets are displayed separately; activity bars offer time-window and per-engine details.

Quick start

npm

npx -y search-rotation --http --open

Requires Node.js 20.3+. For MCP over stdio, configure your client to run npx -y search-rotation with no arguments. The server is also listed in the official MCP Registry as io.github.robin-bially/search-rotation.

Homebrew

brew install robin-bially/tap/search-rotation
search-rotation --http --open

Homebrew installs the required Node.js runtime. For MCP over stdio, configure your client to run search-rotation with no arguments. See the Homebrew tap for upgrades and details.

Pinned release from GitHub

npx -y --allow-git=all github:robin-bially/search-rotation-mcp#v0.5.0

Use this when you want a fixed version instead of the current npm release.

Add your provider keys in the dashboard, then connect your assistant using the MCP client setup guide for Codex, Claude, Cursor, or OpenCode.

What you get

• Search and fetch: independent rotation for web search and Markdown page extraction.
• Automatic failover: quota-aware ordering, rate-limit cooldowns, and request timeouts.
• Local dashboard: API keys, drag-and-drop engine order, per-engine calls and errors, separate quota balances, and interactive request history.
• Local or remote: MCP over stdio or authenticated Streamable HTTP.

Providers: Tavily · Firecrawl · Parallel · Exa · Google PSE · Jina Reader · DuckDuckGo HTML. Keyless access and quota accounting vary by provider.

MCP tools

search_web · fetch_url · get_engine_status · update_engine_config · open_dashboard

Every name pairs an action with its target. update_engine_config covers what the dashboard's engine list does: enable and disable engines, set the search and fetch rotation order, and set monthly or daily limits per engine. API keys stay in the dashboard. If your client keeps an allowlist of tool names, replace the earlier web_search and engine_status entries there.

The dashboard's MCP Tools tab explains each tool, its parameters, and copyable example calls.

Actual dashboard UI, shown in English.

Search time filters

search_web accepts the following arguments:

Use either timeRange or explicit dates. One-sided bounds and equal start/end dates are allowed; invalid calendar dates, reversed bounds, and mixing relative and explicit filters are rejected before any provider request.

{"name":"search_web","arguments":{"query":"AI inference research","timeRange":"week","numResults":5}}

{"name":"search_web","arguments":{"query":"AI inference research","startDate":"2026-08-01","endDate":"2026-08-31"}}

Relative windows are resolved once per request into UTC dates, including across failover. These are date filters, not exact rolling 24-hour windows; month and year mean 30 and 365 days, not calendar arithmetic. Exa receives the start of the first UTC day and the end of the last UTC day.

Provider support and rotation

This table describes implemented support in search-rotation, not every upstream feature. Tavily receives start_date / end_date, Firecrawl receives a custom tbs range, and Exa receives startPublishedDate / endPublishedDate. Firecrawl is conservatively excluded for one-sided dates; the other skipped paths have no implemented date mapping.

The router excludes incompatible providers before quota checks and rotation, including when an incompatible engine is preferred. Each eligible provider pool has its own rotation cursor, so mixing filtered and unfiltered searches does not starve providers. Existing quota priorities, cooldowns, strict-free rules, and failover apply within that pool. If no compatible provider is available, the search fails explicitly; it never retries without the filter. Searches without time arguments retain access to all otherwise eligible providers.

Results show a publication date when supplied by the adapter, and request history includes the requested period. Provider date metadata can be estimated or missing: Tavily filters publication or update dates, Exa filters estimated publication dates, and Firecrawl uses its search index's date interpretation. Exact boundary inclusion follows the provider; this is not an independent verification of each page's publication date. See the Tavily, Firecrawl, and Exa references.

After updating, reconnect your MCP client to load the new tool schema.

Releasing

VERSION=0.4.10 ./scripts/release.sh --publish

The script checks the package (npm ci, build, tests, npm run smoke:package), bumps the version, points the document pins at it, commits and tags, packs the tarball with its checksum, creates the GitHub release, updates the formula in robin-bially/homebrew-tap and finally waits until the npm registry serves the new version. Without --publish it only prepares the artifacts in .build/releases; --dry-run checks the prerequisites, --draft creates a draft release and --force tolerates a dirty tree. The tap is cloned temporarily when TAP_DIR is not set, so a fresh checkout is enough.

Publishing to npm happens in .github/workflows/publish.yml through trusted publishing once the GitHub release is published, so the script itself needs no npm credentials. The workflow requires a trusted publisher for search-rotation on npmjs.com that points at this repository and publish.yml.

The same workflow registers the version in the official MCP Registry from server.json. Distribution & listings tracks every channel, who keeps it current, and what each one still needs.

Learn more

Client setup · Operations & configuration (DE) · Distribution & listings · Releases · CI · MIT license