# SearchPipe

> SearchPipe is a web search API and MCP server for developers building AI agents. One call runs multi-engine retrieval → full-text fetching → LLM reranking → optional summarization, returning structured results with relevance scores. A Tavily alternative.
>
> Every result ships with verifiable evidence — span-level quotes, claim verdicts, and source authority — not just links.

## Core capabilities

- Multi-engine retrieval: Aggregated results from multiple search engines on self-hosted infrastructure.
- Full-text extraction: Fetches top-N pages and extracts the article body, stripping navigation and ads.
- LLM relevance reranking: Every result scored 0–1 against the query intent and sorted by relevance.
- AI answers with citations: Optional summary with citation markers, flagged ai_generated=true.
- Evidence spans: Every result carries a verbatim supporting quote with character offsets from the fetched page text.
- Claim verdicts: Audited answers break into per-statement verdicts — supported, partial, unsupported or inconclusive — with an overall audit score.
- URL health: Each evidence item reports fetch status code, redirects and the final URL of the cited page.
- Source authority: Sources are classified into six tiers: official, press, academic, preprint, community, unknown.
- Remote MCP server: streamable-http MCP endpoint; tool name searchpipe_search, key in the URL.
- REST API: POST /search callable from any language, framework or backend service.
- Automatic refunds: Failed calls (502 upstream failure, 400 output violation) are refunded idempotently.
- Rate limiting: Redis sliding window: 100 requests per minute with a burst of 20 by default.
- Result caching: Identical query and parameters served from cache for 300 seconds by default.

## How SearchPipe differs from Tavily, Exa and Brave

- One call runs the whole pipeline: multi-engine retrieval, full-text extraction, LLM relevance scoring (0–1) and an optional cited answer are all inside POST /search — on Tavily, Exa and Brave, extraction, reranking or answers are separate endpoints billed on top of the search call
- Tavily-compatible interface: same core parameter names and a compatible response shape, so switching from Tavily is a base-URL and API key change
- One-line MCP setup: hosted remote MCP server with the sp- API key embedded in the URL (Brave's MCP server is open source and runs locally)
- Prepaid credits that never expire (subscription credits 30 days); failed calls are auto-refunded; cached calls are still charged at the selected search depth
- Full side-by-side matrix: [Compare web search APIs](https://searchpipe.tech/compare) — competitor data verified 2026-09 against official pricing pages

## Verification API

The verification endpoint checks whether live web evidence actually supports a claim.
Every result ships with verifiable evidence — span-level quotes, claim verdicts, and source authority — not just links.
Verdicts come back on four labels (supported, partial, unsupported, inconclusive) with machine-readable reason codes.

- Endpoint: `POST https://searchpipe.tech/verify` with body `{"claim": "..."}` · Auth: same sp- API key as /search
- Status: Generally available — billing is per request, an `inconclusive` verdict is a valid deliverable, not a refundable failure, and a deployment with the endpoint switched off answers 404
- Evidence-native results: every /search result carries an evidence object — a span-level quote with character offsets, a source-authority tier (official, press, academic, preprint, community, unknown) and fetch metadata
- Audited answers: an advanced search with include_answer=true returns answer_audit — every statement in the generated answer is checked against its cited source with a four-label verdict, plus an overall audit score
- Audit your own pipeline: pass your own cited URLs via evidence_urls to check any LLM answer citations before publishing
- Reason codes (template-locked): doi_match, title_match, quote_span_found, no_supporting_evidence, fetch_failed, paywall, low_authority, conflicting_evidence
- Not a link checker: URL health (status code, redirects, Wayback snapshots) is one field per evidence item; the verdict comes from span-level quote matching and source authority
- Product page: [Citation & Claim Verification API](https://searchpipe.tech/verify-api)

## MCP integration

- Remote MCP server (streamable-http): `https://searchpipe.tech/mcp/?api_key=sp-YOUR_API_KEY` — the API key travels in the URL, no local process to run
- Tools: `searchpipe_search` (web search), `searchpipe_verify` (claim verification), `searchpipe_account_balance` (remaining credits, free to call)
- One-line setup (Claude Code): `claude mcp add --transport http searchpipe "https://searchpipe.tech/mcp/?api_key=sp-YOUR_API_KEY"`
- Per-client setup guides: [MCP integration](https://searchpipe.tech/mcp-server) · [Agent setup guide](https://searchpipe.tech/agents) (Claude Code, Cursor, Codex, OpenCode, Pi, custom MCP agents)
- Agent self-onboarding: [SKILL.md](https://searchpipe.tech/agent-setup/SKILL.md) — a one-page doc AI agents read to configure themselves
- Not using MCP? Call `POST https://searchpipe.tech/search` directly with the same key: [Docs](https://searchpipe.tech/docs)

## API quick facts

- Access: two first-class paths, equal priority — remote MCP Server (streamable-http, key can be embedded in the URL) and REST API `POST /search`; same pipeline, parameters and billing
- Endpoint: `POST https://searchpipe.tech/search` · Auth: `Authorization: Bearer sp-YOUR_KEY` or `X-API-Key: sp-YOUR_KEY`
- Account balance: `GET https://searchpipe.tech/v1/account/balance` with the same sp- API key returns remaining credits (balance, permanent, expiring, upcoming, next_expiry) — free to call, no usage logged; the same data is available over MCP via the `searchpipe_account_balance` tool
- Versioning: /v1/ path prefix available for all REST endpoints (POST /v1/search, /v1/verify, /v1/verify/batch, /v1/extract); existing paths remain valid aliases
- Official SDKs: Python "pip install searchpipe" (PyPI) / TypeScript source in the integration repo, npm package pending (wrapper over POST /search)
- Request parameters: `query` (required, ≤2000 chars) · `max_results` (int, default 5, range 1–20; up to that many results are returned) · `include_answer` (bool, default false; deprecated alias of output_type='answer') · `include_raw_content` (bool, default false) · `search_depth` (`basic` = 1 credit, `advanced` = 2 credits, `fast` = 0.5 credits with all LLM steps skipped) · `include_domains` / `exclude_domains` (array of string; a bare domain matches the domain and any subdomain, a `*.` prefix matches subdomains only, `domain/path` also requires the URL path prefix; exclude wins) · `time_range` (`day`/`week`/`month`/`year`) · `start_date` / `end_date` (ISO date; results without a known publish date are not filtered) · `topic` (`general`/`news`) · `language` / `country` (upstream market locale) · `safe_search` (0/1/2) · `include_images` (bool) · `output_type` (`search_results`/`answer`/`structured`) + `output_schema` (flat JSON Schema) · `objective` (agent intent, feeds reranking) · `stream` (bool, SSE)
- Response fields: `query` · `answer` (when answer mode) · `output` (structured JSON per output_schema, when output_type='structured') · `ai_generated` · `results[]` with `url`, `title`, `content`, `score` (LLM 0–1 on basic/advanced; unnormalized engine score on fast) and `raw_content` · `evidence` per result (quote, span offsets, source tier, fetch status — when evidence is enabled) · `answer_audit` on advanced answers (per-statement verdicts plus score, when answer audit is enabled) · `images[]` (when include_images=true) · `warnings` / `notices` (observability channels) · `request_id` · `response_time` · `usage` (`credits_charged`, `cached`) · per-result `id` and `published_date` — all additive, older fields unchanged
- Streaming: `"stream": true` (answer modes only) responds with text/event-stream — meta → results → delta* → done (full response envelope); stream-started failures arrive as an error event with the call refunded
- Error codes: `400` content violation (auto-refunded) · `401` invalid key · `402` insufficient credits · `429` rate limited (see `Retry-After`) · `502` upstream failure (auto-refunded)
- Pricing: credit-based; basic search 1 credit, advanced search 2 credits, fast search 0.5 credits; $0.005 = 1 credit (bulk recharges from $0.0040); recharges from $40 never expire; subscriptions Starter $10/2,500 · Pro $20/5,000 · Max $35/10,000 credits per month (valid 30 days)
- Free tier: 1,000 free credits per UTC calendar month, renewed on the first verified-account sign-in of each month, with full API and MCP capabilities; failed requests are auto-refunded
- Result cache: identical query and parameters hit the cache for 300 seconds by default
- Rate limit: sliding window, 100 requests/minute with burst 20 by default
- Company: SearchPipe (founded 2026) · contact: privacy@searchpipe.tech

## Key pages

- [Home](https://searchpipe.tech/): product positioning, capabilities and live demo
- [Web Search API for AI Agents](https://searchpipe.tech/web-search-api-for-ai-agents): what agents need from web search and how SearchPipe delivers it
- [Web Search MCP](https://searchpipe.tech/web-search-mcp): what is a Web Search MCP server and how to connect any MCP-compatible agent
- [Docs](https://searchpipe.tech/docs): documentation home — quickstart, authentication, search modes, errors, limits and SDKs
- [Search API reference](https://searchpipe.tech/docs/search): POST /v1/search — every parameter, response field, evidence and answer audit, structured output, streaming
- [Extract API reference](https://searchpipe.tech/docs/extract): POST /v1/extract — batch URL extraction with per-URL failures and automatic refunds
- [Verify API reference](https://searchpipe.tech/docs/verify): POST /v1/verify and /v1/verify/batch — four-label verdicts, reason codes, request and response fields
- [MCP server reference](https://searchpipe.tech/docs/mcp): remote MCP tools, connection URL, authentication and client configuration
- [Search API](https://searchpipe.tech/search-api): Web search API for AI agents and applications — quickstart, parameters, response schema and use cases
- [Extraction API](https://searchpipe.tech/extract-api): fetch readable text from 1-20 URLs in one call, per-URL errors and automatic failure refunds
- [Verification API](https://searchpipe.tech/verify-api): claim verification product page — audit your own pipeline with four-label verdicts and span-level evidence
- [MCP integration](https://searchpipe.tech/mcp-server): connect any MCP-compatible AI agent, setup and troubleshooting
- [Agent setup guide](https://searchpipe.tech/agents): per-client MCP config snippets (Claude Code, Cursor, Codex, OpenCode, Pi) plus custom MCP agents
- [Integrations](https://searchpipe.tech/integrations): MCP server, Python and TypeScript SDKs, agent self-setup and MCP directories
- [Blog](https://searchpipe.tech/blog): engineering notes, including how the public benchmark run was built
- [Benchmark scorecard, honestly](https://searchpipe.tech/blog/search-api-benchmarks-2026): how the 300-question public benchmark run was built and what the numbers do and do not mean
- [Pricing](https://searchpipe.tech/pricing): $0.005 = 1 credit (bulk recharges from $0.0040); recharges never expire, subscription credits valid 30 days
- [Compare web search APIs](https://searchpipe.tech/compare): SearchPipe vs Tavily vs Exa vs Brave — retrieval, pipeline, integration and pricing side by side (competitor data verified 2026-09)
- [Web search API benchmarks](https://searchpipe.tech/benchmarks): SearchPipe-run same-question comparisons, accuracy confidence intervals, observed latency and public list prices; no imported vendor scores
- [FAQ](https://searchpipe.tech/faq): billing, rate limits, key safety and more
- [Contact](https://searchpipe.tech/contact): reach support by email for payment and sign-in issues
- [Terms of Service](https://searchpipe.tech/terms): credits, refunds and acceptable use
- [Privacy Policy](https://searchpipe.tech/privacy): data collection, processors and your rights
- [Agent self-setup guide](https://searchpipe.tech/agent-setup/SKILL.md): a one-page onboarding doc for AI agents

## Docs in Markdown

- [Docs hub](https://searchpipe.tech/docs.md): the Introduction page as plain Markdown
- [Search API reference](https://searchpipe.tech/docs/search.md): POST /v1/search parameters, response fields, errors — as Markdown tables
- [Extract API reference](https://searchpipe.tech/docs/extract.md): POST /v1/extract contract — as Markdown
- [Verify API reference](https://searchpipe.tech/docs/verify.md): POST /v1/verify and /v1/verify/batch contract — as Markdown
- [MCP server reference](https://searchpipe.tech/docs/mcp.md): tools, connection URL, clients — as Markdown