local-web-search 0.1.6

Agent-friendly local web search using SearXNG for snippets and Crawl4AI for full-text retrieval.

Local Web Search

Local Web Search gives AI agents a local web-search backend with two familiar tools:

• web_search asks your local SearXNG instance for search results and returns compact snippets.
• web_fetch uses Crawl4AI to fetch full page text only when the agent asks for a specific result.

Quick Start: OpenAI Agents SDK

Install Local Web Search:

pip install local-web-search

Then install the OpenAI Agents SDK yourself:

pip install openai-agents

Or install both with the convenience extra:

pip install "local-web-search[agents]"

To add repo-local development guidance for agentic coding tools, run this from the project that will use Local Web Search:

local-web-search skill load

The command writes .agents/skills/local-web-search with usage guidance for the OpenAI Agents SDK, LiteLLM-style model adapters, Docker, SearXNG, and Crawl4AI.

[!WARNING] Docker Engine must be running. With build_container_if_missing=True, Local Web Search starts the bundled SearXNG container automatically when it is missing or stopped.

Use the ready-made web_search and web_fetch tools in an agent:

from agents import Agent, Runner
from local_agentic_search.agent_tools import build_agent_tools

web_search, web_fetch = build_agent_tools(build_container_if_missing=True)

agent = Agent(
    name="Research assistant",
    instructions=(
        "Use web_search for current web information. Search results are "
        "snippets. Call web_fetch with a result_id before relying on page "
        "details not present in a snippet."
    ),
    model="gpt-4.1-mini",
    tools=[web_search, web_fetch],
)

result = Runner.run_sync(agent, "Find recent information about Crawl4AI.")
print(result.final_output)

That is the main path: install the package, install the Agents SDK directly or through the extra, keep Docker Engine running, and let build_agent_tools() do the SearXNG startup.

What It Does

Local Web Search keeps web discovery local. Your agent searches first, gets small result snippets with stable result_id values, then fetches full page text only for the result it actually needs. SQLite caching keeps result IDs and page text stable across calls, so an agent can search, reason, fetch, and continue without losing track of sources.

flowchart LR
    Agent["Agent or app"] --> Search["web_search(query)"]
    Search --> Service["Local Web Search"]
    Service --> Cache[("SQLite cache")]
    Service --> SearXNG["Local SearXNG"]
    SearXNG --> Results["Snippets + result_id"]
    Results --> Agent
    Agent --> Fetch["web_fetch(result_id)"]
    Fetch --> Service
    Service --> Crawl4AI["Crawl4AI"]
    Crawl4AI --> Page["Full page text slice"]
    Page --> Agent

Why this shape works:

• Agents see normal tool names: web_search for discovery and web_fetch for source text.
• Search results stay compact, which keeps context windows cleaner.
• Full page fetches are deliberate, cached, and bounded by character limits.
• Execution happens in your local environment, pointed at your local SearXNG service.

Install

Install the package:

pip install local-web-search

Then point it at a SearXNG instance. If you cloned this repository, the bundled Docker Compose setup can start one for you:

docker compose up -d searxng
local-web-search doctor

If you already run SearXNG somewhere else, set SEARXNG_BASE_URL instead:

$env:SEARXNG_BASE_URL = "http://127.0.0.1:8888"
local-web-search doctor

The Python import package is local_agentic_search:

from local_agentic_search import LocalSearchService

Optional Integrations

Most users do not need extras. The base package is enough for the CLI, direct Python usage, and custom tool loops:

pip install local-web-search

If you want the OpenAI Agents SDK path, you can either install the SDK manually:

pip install local-web-search
pip install openai-agents

Or use the extra as a shortcut:

pip install "local-web-search[agents]"

What each extra adds:

Install	Adds	Use it when	Why it is normally not needed
local-web-search	CLI, Python service, SearXNG client, Crawl4AI fetch support	You want local search/fetch from the CLI or your own Python code	This is the normal path
local-web-search[server]	FastAPI and Uvicorn	You want local-web-search serve and HTTP routes	The CLI and direct Python imports do not need an HTTP server
local-web-search[agents]	OpenAI Agents SDK	You want ready-made web_search and web_fetch tools for an Agent	You can also install openai-agents yourself
local-web-search[litellm]	OpenAI Agents SDK with LiteLLM adapter dependencies	You want the Agents SDK helper tools with a LiteLLM-backed model	OpenAI-only workflows do not need LiteLLM
local-web-search[server,agents]	Both integration layers	You want both the HTTP API and OpenAI Agents SDK helper tools	Most projects use one integration layer at a time

The integrations are intentionally separate because the base package already does the core job. The extras keep optional framework dependencies out of your environment until you actually need them.

Quick Start From This Repo

Clone the repository if you want the bundled Docker Compose SearXNG setup:

git clone https://github.com/maestromaximo/local-web-search.git
cd local-web-search
docker compose up -d searxng
pip install -e ".[server,agents,dev]"
crawl4ai-setup

Check that SearXNG is reachable:

local-web-search doctor

Search from the CLI:

local-web-search search "OpenAI Agents SDK function tools" --max-results 5

Fetch full text for a result returned by search:

local-web-search fetch res_... --max-chars 4000

Fetch a URL directly:

local-web-search fetch https://example.com --max-chars 4000

Run the HTTP API:

local-web-search serve --host 127.0.0.1 --port 8099

Install the repo-local development skill:

local-web-search skill load

CLI

local-web-search doctor
local-web-search search "query" [--max-results 5] [--language en]
local-web-search fetch res_... [--start 0] [--max-chars 4000]
local-web-search fetch https://example.com [--max-chars 4000]
local-web-search serve [--host 127.0.0.1] [--port 8099]
local-web-search skill load [--project-dir .] [--force]

Run local-web-search --help or local-web-search <command> --help for the full command reference.

OpenAI Agents SDK Usage

Install Local Web Search plus the OpenAI Agents SDK:

pip install local-web-search
pip install openai-agents

Or use the convenience extra:

pip install "local-web-search[agents]"

With Docker Engine running, build tools named web_search and web_fetch:

from agents import Agent, Runner
from local_agentic_search.agent_tools import build_agent_tools

web_search, web_fetch = build_agent_tools(build_container_if_missing=True)

agent = Agent(
    name="Research assistant",
    instructions=(
        "Use web_search when current web information is useful. Search results "
        "are snippets. Call web_fetch with a result_id before relying on page "
        "details not present in a snippet."
    ),
    model="gpt-4.1-mini",
    tools=[web_search, web_fetch],
)

result = Runner.run_sync(agent, "Find recent information about Crawl4AI.")
print(result.final_output)

By default, build_agent_tools() assumes SearXNG is already running and prints a yellow console warning once per process. To let the tool factory start SearXNG when the container is missing or stopped:

web_search, web_fetch = build_agent_tools(build_container_if_missing=True)

If port 8888 is already used on the host, choose another host port and keep Docker plus the Python client aligned:

web_search, web_fetch = build_agent_tools(
    build_container_if_missing=True,
    searxng_port=8899,
)

If the SearXNG container is already running on the old port, recreate it after changing ports:

docker compose down
$env:LOCAL_WEB_SEARCH_SEARXNG_PORT = "8899"
docker compose up -d searxng

To silence the warning while keeping startup manual:

web_search, web_fetch = build_agent_tools(suppress_docker_warning=True)

LiteLLM or special model adapters

The Local Web Search tools are local OpenAI Agents SDK function tools, so they can stay attached to the agent while you change the model adapter:

pip install "local-web-search[litellm]"

from agents import Agent, Runner
from agents.extensions.models.litellm_model import LitellmModel
from local_agentic_search.agent_tools import build_agent_tools

web_search, web_fetch = build_agent_tools(build_container_if_missing=True)

agent = Agent(
    name="Research assistant",
    instructions=(
        "Use web_search for current web information. Search results are "
        "snippets. Call web_fetch with a result_id before relying on page "
        "details not present in a snippet."
    ),
    model=LitellmModel(model="openai/gpt-4.1-mini"),
    tools=[web_search, web_fetch],
)

result = Runner.run_sync(agent, "Find recent information about SearXNG.")
print(result.final_output)

Validate tool calling, streaming, usage reporting, and tracing against the exact provider behind LiteLLM before relying on it in production.

Responses API Tool Schemas

You do not need the agents extra for direct OpenAI API tool loops. Use the schema helper:

from local_agentic_search.tool_schemas import responses_tool_schemas

tools = responses_tool_schemas()

Your application still executes web_search and web_fetch locally and returns their JSON outputs as function call outputs.

HTTP API

Install the server integration only for this path:

pip install "local-web-search[server]"

Start the server with:

local-web-search serve

Routes:

• GET /health
• GET /search?q=...&max_results=5
• POST /search
• POST /fetch
• GET /openai/tools

See examples/http_client_example.py for a minimal async HTTP client.

Examples

• examples/agents_sdk_example.py: build OpenAI Agents SDK tools named web_search and web_fetch.
• examples/http_client_example.py: call the local HTTP API with httpx.
• examples/responses_schema_example.py: print the Responses API tool schemas.

Star Tracker

If Local Web Search saves you from wiring search tools by hand, starring the repo helps track interest and keeps the project visible:

Search Result Shape

Each web_search result includes:

{
  "result_id": "res_...",
  "search_id": "search_...",
  "position": 1,
  "title": "Page title",
  "url": "https://example.com",
  "snippet": "Compact SearXNG snippet",
  "site_links": [],
  "full_text_available": true,
  "full_text_command": {
    "tool": "web_fetch",
    "arguments": {
      "result_id": "res_...",
      "start": 0,
      "max_chars": 4000
    }
  },
  "full_text_command_text": "web_fetch(result_id='res_...', start=0, max_chars=4000)"
}

web_fetch returns a page slice with start, end, total_chars, has_more, and a next_fetch_command when more text is available.

Configuration

Environment variables:

• SEARXNG_BASE_URL: defaults to http://127.0.0.1:8888
• LOCAL_WEB_SEARCH_CACHE: defaults to .cache/local_agentic_search.sqlite3
• LOCAL_WEB_SEARCH_RESULTS_TTL_SECONDS: defaults to 86400
• LOCAL_WEB_SEARCH_PAGES_TTL_SECONDS: defaults to 604800
• LOCAL_WEB_SEARCH_FETCH_CHARS: default fetch slice size, defaults to 4000
• LOCAL_WEB_SEARCH_MAX_FETCH_CHARS: maximum fetch slice size, defaults to 20000
• LOCAL_WEB_SEARCH_CRAWL_TIMEOUT_MS: Crawl4AI timeout, defaults to 45000
• LOCAL_WEB_SEARCH_DOCKER_COMPOSE_FILE: optional compose file path
• LOCAL_WEB_SEARCH_DOCKER_CONTAINER: optional SearXNG container name
• LOCAL_WEB_SEARCH_SEARXNG_HOST: host used by the Python client when SEARXNG_BASE_URL is unset, defaults to 127.0.0.1
• LOCAL_WEB_SEARCH_SEARXNG_PORT: host port for the bundled SearXNG service, defaults to 8888
• LOCAL_WEB_SEARCH_SEARXNG_BIND: Docker bind address for the bundled SearXNG service, defaults to 127.0.0.1

The automatic Docker path uses a fast docker container inspect check. If the configured container is paused, it runs docker container unpause; if it is missing or stopped, it runs docker compose up -d --build searxng.

License and Notices

Local Web Search is licensed under the Apache License 2.0.

SearXNG is a separate service licensed under the GNU Affero General Public License v3.0 or later. Crawl4AI is licensed under the Apache License 2.0. See THIRD_PARTY_NOTICES.md for details.

You are responsible for respecting robots.txt, website terms, copyright, authentication boundaries, privacy rules, and rate limits when searching and fetching public web content.