SearchForge

<div align="center"> # SearchForge ### Open-source web search API and MCP server for LLMs, agents, and RAG. [](https://github.com/divyanshu-iitian/SearchForge/actions/workflows/ci.yml) [](LICENSE) [](package.json) [](https://modelcontextprotocol.io) [](https://divyanshu-iitian.github.io/SearchForge/) [](https://registry.modelcontextprotocol.io/v0.1/servers?search=io.github.divyanshu-iitian/searchforge) **One local gateway. Four search capabilities. Clean Markdown. REST, MCP, CLI, and TypeScript.** [Website](https://divyanshu-iitian.github.io/SearchForge/) · [Quick start](#quick-start) · [Free tools](#free-tools) · [MCP](#mcp) · [API](#rest-api) · [Design](#how-it-works) </div> --- **SearchForge is a free, open-source web search API and MCP server for LLMs, AI agents, and retrieval-augmented generation (RAG) pipelines.** It provides a predictable retrieval layer without forcing every project to integrate a paid search vendor. SearchForge routes each query to the right source, isolates provider failures, deduplicates URLs, fuses rankings, and turns public pages into LLM-ready Markdown. It does **not** generate answers, hide citations, scrape public SearXNG instances, or send telemetry. ## What you get | Capability | Default source | Cost / credentials | | ----------- | ----------------------------------- | ------------------------------ | | `web` | Wikipedia; optional private SearXNG | No key / self-hosted | | `code` | GitHub repository search | No key; token optional | | `academic` | Crossref works and DOI metadata | No key | | `community` | Hacker News via Algolia | No key, community service | | `read_url` | Jina Reader | No key, currently rate-limited | SearchForge starts with all no-key adapters enabled. A GitHub token only raises the public API quota, and Brave remains an optional keyed backend. Broad, independent web metasearch is provided by the included SearXNG stack. ## Quick start ### Zero-key local CLI ```bash git clone https://github.com/divyanshu-iitian/SearchForge.git cd SearchForge npm install npm run build node dist/cli.js search "open source agent frameworks" --category code node dist/cli.js search "retrieval augmented generation" --category academic node dist/cli.js search "local LLM tooling" --category community node dist/cli.js read "https://example.com" node dist/cli.js doctor ``` ### Full web search with private SearXNG ```bash docker compose up --build ``` ```bash curl -s http://localhost:3000/v1/search \ -H "content-type: application/json" \ -d '{"query":"open source vector databases","category":"web","limit":5}' ``` This starts SearchForge on port `3000` and a private, JSON-enabled SearXNG on port `8080`. Before exposing the stack, change the SearXNG secret, set `SEARCHFORGE_API_KEY`, and terminate TLS at a trusted proxy. ## Free tools ### Search by capability ```bash searchforge search "browser agent" --category code searchforge search "semantic reranking" --category academic --json searchforge search "Show HN search engine" --category community ``` Categories prevent irrelevant providers from being queried. An explicit `providers` list overrides category routing, which is useful for evaluations. ### Read a URL as Markdown ```bash searchforge read "https://example.com/article" ``` `read_url` accepts public HTTP(S) URLs only. Credentials, localhost, private IP literals, and non-web protocols are rejected. Responses are size-bounded, timed out, and cached. ### Diagnose the whole retrieval path ```bash searchforge doctor ``` Doctor performs real, bounded probes and reports each provider's access tier, capability, latency, and error. A failed source produces `degraded`, not a misleading all-or-nothing status. ## MCP SearchForge exposes three stdio tools: - `web_search` — routed, citation-ready structured search - `read_url` — clean Markdown from a public URL - `search_status` — live capability and latency report ```json { "mcpServers": { "searchforge": { "command": "node", "args": ["/absolute/path/to/SearchForge/dist/mcp.js"], "env": { "SEARCHFORGE_SEARXNG_URL": "http://localhost:8080" } } } } ``` The search and status tools return MCP structured content as well as readable text. SearchForge is also published in the [official MCP Registry](https://registry.modelcontextprotocol.io/v0.1/servers?search=io.github.divyanshu-iitian/searchforge) as `io.github.divyanshu-iitian/searchforge`. To run the registry-backed OCI image directly from any MCP client that supports a Docker command: ```json { "mcpServers": { "searchforge": { "command": "docker", "args": [ "run", "--rm", "-i", "ghcr.io/divyanshu-iitian/searchforge-mcp:0.2.0" ] } } } ``` ## REST API ### Search ```http POST /v1/search Content-Type: application/json { "query": "open source reranking models", "category": "academic", "limit": 8, "language": "en", "freshness": "month", "safeSearch": "moderate" } ``` ```json { "schemaVersion": "1.0", "query": "open source reranking models", "category": "academic", "results": [ { "title": "Example work", "url": "https://doi.org/10.0000/example", "snippet": "Authors · Publisher · journal-article", "source": "crossref", "sources": ["crossref"], "score": 0.016393 } ], "providers": [ { "provider": "crossref", "ok": true, "latencyMs": 241, "resultCount": 8 } ], "tookMs": 243, "cached": false } ``` ### Read ```http POST /v1/read Content-Type: application/json {"url":"https://example.com/article"} ``` Other endpoints: ```text GET /healthz Process liveness GET /v1/providers Configured capabilities and access tiers GET /v1/doctor Live dependency health ``` See the full [OpenAPI contract](openapi.yaml). ## TypeScript SDK ```ts import { CrossrefProvider, GithubProvider, JinaReader, SearchForge, } from "searchforge-rag"; const forge = new SearchForge({ providers: [new GithubProvider(), new CrossrefProvider()], reader: new JinaReader(), timeoutMs: 8_000, }); const evidence = await forge.search({ query: "agentic retrieval", category: "academic", limit: 10, }); const page = await forge.read("https://example.com/research"); ``` Until an npm release is published: ```bash npm install github:divyanshu-iitian/SearchForge ``` ## Provider details | Provider | Capability | Access | Enabled | | ------------------------------------------------------------------------------ | ---------------------- | ------------------------------------------------------------------------ | ------------------------- | | [SearXNG](https://docs.searxng.org/) | Web | Self-hosted, no vendor fee | `SEARCHFORGE_SEARXNG_URL` | | Wikipedia | Web knowledge fallback | No key | Always | | [GitHub](https://docs.github.com/rest/search/search) | Code repositories | No key; 60 unauthenticated REST requests/hour, search has tighter limits | Always | | [Crossref](https://www.crossref.org/documentation/retrieve-metadata/rest-api/) | Academic metadata | No key; `mailto` recommended | Always | | [HN Algolia](https://hn.algolia.com/api) | Community | No key; community-operated availability | Always | | [Jina Reader](https://jina.ai/reader/) | URL to Markdown | No key; documented no-key quota currently 20 RPM | Always | | [Brave Search](https://brave.com/search/api/) | Web | API key | `BRAVE_SEARCH_API_KEY` | SearchForge intentionally does not configure public SearXNG instances. They often disable JSON or limit automated traffic; the Docker stack is the stable free path. ## How it works ```text Agent / RAG / MCP client | validate + route | +--------+---------+-----------+ | | | | web code academic community read_url | | | | | SearXNG GitHub Crossref Hacker News Jina Reader Wikipedia +--------+---------+-----------+ | normalize -> canonicalize -> deduplicate -> reciprocal-rank fusion | versioned evidence + provenance + per-source health ``` Each idempotent provider call has its own abortable timeout. One outage cannot erase healthy results. Tracking parameters are removed before deduplication, and every contributing provider remains in `sources`. This capability-first design is inspired by [Agent Reach](https://github.com/Panniantong/Agent-Reach). Agent Reach helps an agent operate many upstream tools directly; SearchForge complements that approach with one stable, embeddable retrieval API for RAG applications. ## Configuration | Variable | Default | Purpose | | ------------------------------- | ----------: | ------------------------------ | | `SEARCHFORGE_SEARXNG_URL` | unset | Private SearXNG base URL | | `GITHUB_TOKEN` | unset | Optional GitHub quota increase | | `CROSSREF_MAILTO` | unset | Crossref polite-pool identity | | `BRAVE_SEARCH_API_KEY` | unset | Optional Brave backend | | `SEARCHFORGE_API_KEY` | unset | REST bearer or `x-api-key` | | `SEARCHFORGE_PORT` | `3000` | REST port | | `SEARCHFORGE_HOST` | `127.0.0.1` | Bind address | | `SEARCHFORGE_TIMEOUT_MS` | `8000` | Per-dependency timeout | | `SEARCHFORGE_CACHE_TTL_MS` | `300000` | In-memory cache TTL | | `SEARCHFORGE_CACHE_MAX_ENTRIES` | `500` | Cache entry bound | | `SEARCHFORGE_RATE_LIMIT` | `60` | Requests/client/minute | ## Production boundary - Set an API key before binding to a public interface. - Search results and page content are untrusted input; delimit them and apply prompt-injection defenses. - The built-in cache and rate limiter are process-local. Use shared infrastructure for multiple replicas. - Provider bodies and credentials are excluded from surfaced errors. - `healthz` proves the process is alive; `/v1/doctor` checks dependencies. See [SECURITY.md](SECURITY.md), [CONTRIBUTING.md](CONTRIBUTING.md), and [CHANGELOG.md](CHANGELOG.md). ## Principles 1. Evidence over generated answers 2. Free and self-hosted paths before vendor lock-in 3. Partial results over total failure 4. Honest capability and quota reporting 5. Stable contracts and explicit provenance 6. No telemetry by default ## License MIT © Divyanshu. If SearchForge helps your agent, star the repository and share your integration in [Discussions](https://github.com/divyanshu-iitian/SearchForge/discussions).