TinySearch

Self-hosted web research for MCP agents.

TinySearch gives local AI agents a web-research tool they can actually use: search the web, rerank results, crawl the best pages, extract the most relevant chunks, and return a source-grounded prompt your LLM can answer from.

TinySearch terminal demo showing a source-grounded research prompt

No hosted dashboard. No account system. No analytics. No scraped-data cache.

Just search -> crawl -> rerank -> grounded prompt.

Why people use it

• Add web research to Cursor, Cline, Roo Code, Claude Desktop, or any MCP client.
• Keep source URLs attached to the evidence your model sees.
• Avoid dumping full webpages into context.
• Run with local ONNX embeddings by default, or bring an OpenAI-compatible embedding API.
• Use SearXNG by default, with a DuckDuckGo HTML fallback when configured.
• Keep the stack small enough to run locally in Docker.

TinySearch is built for local agents, prototypes, personal workflows, and small systems where source-grounded web research matters more than running a full search product.

Quick start

Run TinySearch with its own SearXNG instance as an MCP server over Streamable HTTP. Docker Compose loads the configuration directly from GitHub, so you do not need to clone the repository or create any configuration files:

docker compose -f "https://github.com/MarcellM01/TinySearch.git#main:compose.quickstart.yaml" up -d

Then connect your MCP client to:

{
  "mcpServers": {
    "tinysearch": {
      "url": "http://localhost:8000/mcp"
    }
  }
}

Stop and remove the containers later with:

docker compose -f "https://github.com/MarcellM01/TinySearch.git#main:compose.quickstart.yaml" down

TinySearch exposes three MCP tools:

get_current_datetime()
research(query)
scrape_url(url, query)

Typical routing:

• Use research(query) when the agent needs to discover relevant URLs.
• Use scrape_url(url, query) when the user already provided a URL, or when research found the page to inspect.
• Use get_current_datetime() before time-sensitive research.

The tools return a grounded prompt in the answer field. Your MCP client model uses that prompt to write the final response with citations.

How it works

flowchart TB
    subgraph Row1["Search and choose pages"]
        direction LR
        A[User query] --> B[Web search<br/>SearXNG default, DuckDuckGo fallback]
        B --> C[Filter HTTP results<br/>build title URL domain snippet docs]
        C --> D[Rank search docs<br/>dense + BM25 weighted RRF]
    end

    subgraph Row2["Crawl and build prompt"]
        direction LR
        E[Crawl kept URLs in parallel<br/>crawl4ai markdown] --> F[Truncate and chunk markdown]
        F --> G[Rank combined chunk pool<br/>dense + BM25 weighted RRF]
        G --> H[Dedupe chunks<br/>apply source quotas and fill]
        H --> I[Build source-grounded prompt]
    end

    Row1 --> Row2

TinySearch does not directly answer the question. It returns a structured prompt in the MCP tool's answer field, and your client model uses that prompt to produce the final cited response.

QUESTION
What happened in the latest NFL playoffs?

TODAY
2026-05-15

RESULTS
1. Title
   URL
   Relevant extracted text...

2. Title
   URL
   Relevant extracted text...

INSTRUCTIONS
Answer only from the results. Cite source URLs.

Run from source

Use this path if you want to inspect the code, edit TinySearch, or run it as a local stdio MCP server.

git clone https://github.com/MarcellM01/TinySearch
cd TinySearch

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

MCP clients spawn TinySearch from their config. Add it with absolute paths:

macOS / Linux:

{
  "mcpServers": {
    "tinysearch": {
      "command": "/absolute/path/to/TinySearch/.venv/bin/python",
      "args": [
        "/absolute/path/to/TinySearch/servers/mcp_server.py"
      ]
    }
  }
}

Windows:

{
  "mcpServers": {
    "tinysearch": {
      "command": "C:/absolute/path/to/TinySearch/.venv/Scripts/python.exe",
      "args": [
        "C:/absolute/path/to/TinySearch/servers/mcp_server.py"
      ]
    }
  }
}

Template config files live in mcp_templates/.

The repo also includes agentic_coding_templates/global-rules-recommended.md, a global-rules template for agentic coding tools such as Cline and Roo Code. These rules help coding agents call TinySearch only when web research is actually needed.

The server uses stdio by default, which is what Cursor and similar clients expect when they spawn python .../mcp_server.py. To run with sse or streamable-http, set MCP_TRANSPORT when starting the process. Do not put transport in configs/research_config.json.

Docker

The quick start command runs TinySearch over Streamable HTTP on http://localhost:8000/mcp. Docker pulls marcellm01/tinysearch:latest automatically if the image is not already local.

With MCP_TRANSPORT=streamable-http, the image serves Streamable HTTP on /mcp and SSE on /mcp/sse. GET requests to /mcp without an mcp-session-id are treated as the legacy SSE stream. If a client still cannot connect, try MCP_TRANSPORT=sse alone or the stdio Docker setup below.

Docker image tags

Docker images are published automatically when a version tag or GitHub release is created.

• marcellm01/tinysearch:<version> is published for tags such as v0.1.4.
• marcellm01/tinysearch:latest is updated for stable releases.
• Images are built for both linux/amd64 and linux/arm64.

Persistent models and config

For repeated use, keep downloaded models in a Docker volume and mount your local config. The mounted config can also include blocked_domains to exclude sites from search results:

docker run --rm \
  -p 8000:8000 \
  -v tinysearch-models:/data/models \
  -v "$PWD/configs/research_config.json:/config/research_config.json:ro" \
  -e TINYSEARCH_CONFIG_PATH=/config/research_config.json \
  -e MCP_TRANSPORT=streamable-http \
  -e MCP_HOST=0.0.0.0 \
  marcellm01/tinysearch:latest

Example config entry:

"blocked_domains": ["example.com", "spammy-site.test"]

MCP over stdio

Use this mode for MCP clients that launch tools as local commands instead of connecting to a URL. Replace /absolute/path/to/TinySearch with this repo's absolute path:

{
  "mcpServers": {
    "tinysearch": {
      "command": "docker",
      "args": [
        "run",
        "--rm",
        "-i",
        "-v",
        "tinysearch-models:/data/models",
        "-v",
        "/absolute/path/to/TinySearch/configs/research_config.json:/config/research_config.json:ro",
        "-e",
        "TINYSEARCH_CONFIG_PATH=/config/research_config.json",
        "-e",
        "TINYSEARCH_MODELS_DIR=/data/models",
        "marcellm01/tinysearch:latest"
      ]
    }
  }
}

Edit configs/research_config.json to choose embedding_model (fast, balanced, quality, or a custom Hugging Face ONNX repo id). The named Docker volume keeps downloaded model bundles between launches.

Optional HTTP server

Useful when you want HTTP instead of MCP:

uvicorn servers.fastapi_server:app --reload

Endpoints mirror the MCP tools:

• GET /health
• GET /current_datetime — same as get_current_datetime()
• POST /research — body {"query": "..."}; same as research(query)
• POST /scrape — body {"url": "...", "query": "..."}; same as scrape_url(url, query)

POST /scrape returns the same fields as the MCP tool: answer, content_tokens, answer_tokens, truncated, url, title, and retrieved_at (aware UTC).

Errors return {"detail": {"code", "message"}} with stable codes: invalid_url (400), blocked_url (403), unsupported_document (415), empty_content (422), fetch_failed (502), fetch_timeout (504).

URL safety

/scrape and scrape_url accept arbitrary user-supplied URLs and enforce the following checks before fetching:

• only http and https schemes
• URLs with embedded credentials are rejected
• IP literals and resolved addresses that are loopback, private, link-local, multicast, reserved or unspecified are rejected (DNS rebinding is mitigated by rejecting if any resolved address is non-public, not just one)
• the configured blocked_domains list is applied to both the initial URL and the final URL reported by the crawler after redirects

Crawl4AI does not expose intermediate redirect hops, so the safety check runs on the initial URL and the final URL. If you need stricter handling for redirect chains, run TinySearch behind an egress proxy that enforces your policy.

Configuration

Tune research defaults in configs/research_config.json. Set TINYSEARCH_CONFIG_PATH to load a different JSON config file, which is the recommended Docker override pattern.

Set blocked_domains to a JSON list of domains you do not want TinySearch to return or crawl. Entries match the domain and its subdomains, so example.com also blocks www.example.com and news.example.com. URL-style entries such as https://example.com/path are accepted and normalized to their hostname.

The onnx embedding backend uses local ONNX bundles under models/. Starting the MCP server or FastAPI app downloads the configured embedding_model once from Hugging Face when embedding_backend is onnx.

Built-in local presets:

• fast: onnx-models/all-MiniLM-L6-v2-onnx
• balanced: BAAI/bge-small-en-v1.5
• quality: BAAI/bge-base-en-v1.5

You can also set embedding_model to a custom Hugging Face ONNX repo id. Set TINYSEARCH_MODELS_DIR to move the model cache, or use TINYSEARCH_ONNX_MODEL_DIR when you need to point at one exact bundle directory.

Key settings:

• Search: search_top_k, search_rrf_cutoff, search_dense_weight, search_max_results_to_keep, blocked_domains
• Search backend: search_backend, search_backend_url, search_engines, search_region, search_backend_fallback
• Chunks: chunk_rrf_cutoff, chunk_dense_weight, chunk_max_results_to_keep
• Crawl: crawl_max_chunk_tokens, crawl_overlap_tokens, max_concurrent_crawls
• Embeddings: embedding_backend, embedding_model, embedding_openai_env_file, max_concurrent_embedding_calls
• Tokenizer: encoding_name
• Dense input prefixes: dense_query_prefix, dense_document_prefix
• Trace: trace_path

For embedding_backend openai_compatible, add a .env file at the project root, or set embedding_openai_env_file, with:

OPENAI_BASE_URL=
OPENAI_API_KEY=
OPENAI_EMBEDDING_MODEL=

OPENAI_BASE_URL is optional for api.openai.com. EMBEDDING_MODEL and MODEL_NAME are accepted as aliases for OPENAI_EMBEDDING_MODEL.

The research pipeline requires dense embeddings. It raises if search_dense_weight or chunk_dense_weight is set to 0.

Search backends

TinySearch supports two web-search backends and selects between them from config. The defaults aim at the bundled compose setup: SearXNG runs as a sidecar, with the DuckDuckGo HTML scraper kept as an automatic fallback.

Since v0.2, TinySearch defaults to a SearXNG-compatible backend. The bundled Compose files ship a local SearXNG service so the stack works out of the box, while the DuckDuckGo HTML scraper remains available as a configurable fallback.

Available values for search_backend:

• "searxng" (default): query a SearXNG-compatible JSON endpoint. If the call fails and search_backend_fallback is true, TinySearch falls back to DuckDuckGo. With search_backend_fallback: false the SearXNG error surfaces.
• "duckduckgo": skip SearXNG entirely and use the existing DuckDuckGo HTML scraper. This is the escape hatch that preserves pre-0.2 behavior.
• "auto": try SearXNG, then DuckDuckGo on any backend failure (fallback is implied regardless of search_backend_fallback).

A backend "failure" means a real backend error: network/timeout, non-200 HTTP response, a non-JSON SearXNG body, or a DuckDuckGo CAPTCHA / 403. A legitimate empty result set is not a failure and does not trigger fallback.

Minimal config example:

{
  "search_backend": "searxng",
  "search_backend_url": "http://searxng:8080/search",
  "search_engines": ["google", "bing"],
  "search_region": "us-en",
  "search_backend_fallback": true
}

SearXNG JSON output is required

SearXNG ships with the JSON output format disabled by default. The bundled searxng/settings.yml enables it via:

search:
  formats:
    - html
    - json

If TinySearch reports SearchBackendUnavailable: SearXNG did not return JSON, your SearXNG instance is returning HTML — add json to search.formats and restart it.

Environment overrides

• SEARXNG_URL: overrides search_backend_url for the running process. Useful in Docker so the same image can point at different SearXNG endpoints without rebuilding research_config.json.

Compose setup

The bundled compose.yaml starts a searxng service alongside mcp (and optionally fastapi). The mcp and fastapi services reach SearXNG at http://searxng:8080/search over the internal compose network, and have SEARXNG_URL set automatically.

docker compose up

A minimal searxng/settings.yml is committed at the repo root. Override server.secret_key before exposing the SearXNG instance beyond localhost.

Single-container / from-source

When you run TinySearch standalone (e.g. docker run marcellm01/tinysearch:latest or python servers/mcp_server.py), there is no local SearXNG. With the default config (search_backend: "searxng", search_backend_fallback: true) the SearXNG call fails fast on the short connect timeout and TinySearch transparently falls back to DuckDuckGo.

To keep the pre-0.2 behavior with no SearXNG involvement, set:

{ "search_backend": "duckduckgo" }

When not to use TinySearch

TinySearch is not a replacement for a commercial search API or a persistent crawler. It is probably not the right tool if you need:

• guaranteed search coverage
• large-scale indexing
• long-term page caching
• enterprise observability
• production SLA-backed web search

TinySearch vs...

Community

Join the TinySearch Discord for support, release updates, bug reports, and contributor discussion.

Entrypoints

• pipelines.agentic_research.agentic_run: single-turn search, crawl, ranking, and prompt assembly
• servers.mcp_server: MCP server for agent clients
• servers.fastapi_server: optional HTTP API

Tests

Run the unittest suite:

python -m unittest discover tests

Contact

Using TinySearch or want to build on it?

Email me or reach me on Bluesky.

Privacy notes

TinySearch reads the pages it crawls and returns ranked excerpts to the calling client. It does not include credentials in the repo, and .env / trace output should stay local. If you enable openai_compatible embeddings, your embedding provider receives the text snippets sent for vectorization.

License

Source code in this repository is under the MIT License.

When embedding_backend is onnx, TinySearch may download the selected local ONNX embedding bundle at runtime from Hugging Face. Those weights are separate distributions under their model-card licenses; keep license and attribution notices if you ship or redistribute those files. Optional manual export for fast uses sentence-transformers/all-MiniLM-L6-v2 (Apache-2.0).

See NOTICE for Docker and third-party distribution notes.