Agent Search

AgentSearch

Self-hosted search API for AI agents. Layered content extraction with isolated browser rendering. Optional Tor-anonymized stack. No third-party search API keys, no per-query fees, no vendor lock-in. Optional local bearer auth is supported.

git clone https://github.com/brcrusoe72/agent-search.git
cd agent-search
./scripts/prepare-searxng.sh
docker compose up -d
curl "http://localhost:3939/search?q=distributed+consensus+algorithms"

You now have a deduplicated, multi-engine search API running on :3939.

If you enable auth, pass the token on all endpoints except /health, /live, and /ready:

export AGENT_SEARCH_TOKEN="change-me"
curl -H "Authorization: Bearer $AGENT_SEARCH_TOKEN" \
  "http://localhost:3939/search?q=distributed+consensus+algorithms"

Prefer not to use Docker for the API server?

git clone https://github.com/brcrusoe72/agent-search.git
cd agent-search
./scripts/install-native.sh
./scripts/run-native.sh

Native mode requires Python 3.11+ and a reachable SearXNG instance with JSON output enabled. It stores AgentSearch state in ./data. See Native Install.

Existing Docker installations: follow the upgrade procedure before switching from a root-run image.

Verify

Run the self-contained test suite:

python -m venv .venv
. .venv/bin/activate
pip install -r requirements.txt pytest requests
pip install -e sdk
# MCP has a separate Pydantic requirement; keep it isolated.
python -m venv .venv-mcp
.venv-mcp/bin/python -m pip install -e mcp-server pytest
.venv-mcp/bin/python -m pytest tests/test_mcp_contract.py -q
./scripts/prepare-searxng.sh
pytest tests -q
python -m compileall app adapters mcp-server/agent_search_mcp scripts sdk -q
docker compose -f docker-compose.yml config --quiet
docker compose -f docker-compose.yml -f examples/compose.private.yml config --quiet
docker build -t agent-search-api:test .

Those tests mock SearXNG, so they do not require Docker or a running local service.

Run the optional live localhost check:

AGENTSEARCH_INTEGRATION=1 pytest tests -q

If your local instance requires auth:

AGENT_SEARCH_TOKEN="change-me" AGENTSEARCH_INTEGRATION=1 pytest tests -q

Run the optional Docker smoke tests against a running direct/private stack:

AGENT_SEARCH_TOKEN="change-me" \
AGENTSEARCH_DOCKER_INTEGRATION=1 \
pytest tests/test_live_docker.py -q

What it does

AgentSearch wraps SearXNG with a FastAPI layer that adds the pieces agents usually need on top of raw search: deduplication, cross-engine scoring, content extraction, query expansion, domain trust scoring, prompt injection scrubbing, and failure-pattern analysis.

Standard stack — docker compose up gives you search on :3939.

Private stack — docker compose -f docker-compose.yml -f examples/compose.private.yml up adds an anonymized instance on :3940 that routes all traffic through Tor with Snowflake obfuscation. Encrypted DNS via CoreDNS → Cloudflare DoT. Network-level isolation — the private SearXNG instance runs on an internal-only Docker network with no route to the internet except through Tor.

Search engines

AgentSearch delegates engine support to the connected SearXNG instance. The authoritative list for a running stack is:

curl "http://localhost:3939/engines"

The bundled searxng/settings.example.yml explicitly enables 25 engines, including best-effort Google/Startpage/Yahoo entries plus Brave, Bing, DuckDuckGo, Google Scholar, Semantic Scholar, arXiv, Crossref, OpenAlex, PubMed, Bing News, Reuters, Wikinews, Wikipedia, Wikidata, Hugging Face, Reddit, Hacker News, Stack Overflow, GitHub, Docker Hub, and Lobsters.

Run ./scripts/prepare-searxng.sh to create ignored local runtime files at searxng/settings.yml and searxng/settings.tor.yml with generated SearXNG instance secrets. Do not commit those generated files.

Because SearXNG is configured with use_default_settings: true, your live instance may expose additional enabled engines from the installed SearXNG catalog. Use the engines= query parameter to request specific engines, and use /engines to verify what is available in that deployment.

Strategy modes can also use direct no-key providers for vertical search. These do not require paid search APIs: GitHub repository search, MDN search, Docker Hub search, PyPI package metadata, Wikipedia, Wikidata, Hacker News, arXiv, Crossref, OpenAlex, and Semantic Scholar are called directly when a mode selects them. SearXNG remains the broad-web provider for Google/Bing/Brave/DuckDuckGo-style engines, but Google, Startpage, Yahoo, and Reddit are best-effort explicit sources rather than defaults because they are commonly blocked or empty.

Why not just use SearXNG directly?

SearXNG finds pages. AgentSearch fetches and reads them, scores and deduplicates the results, caches them, scrubs prompt injections, detects paywalls, and falls back through several extraction strategies when the first one fails — in one API call. It also logs fetch outcomes so you can see which strategies and domains are actually working.

Endpoints

Search

Content extraction

The extraction chain escalates through strategies until one succeeds:

1. Direct fetch + smart content selectors
2. Readability scoring (paragraph density vs link density)
3. User-agent rotation (Chrome/Safari/Firefox/Edge signatures)
4. Browser render/extract for JS-rendered target pages
5. Wayback Machine (CDX API → latest snapshot)
6. Google Cache
7. Search-about fallback (find coverage elsewhere)
8. Custom adapters (pluggable Python modules from disk)
9. PDF extraction (pdfplumber)
10. YouTube transcript (yt-dlp)

Every request gets fail-closed SSRF protection, prompt injection detection, paywall detection, and streamed content limits automatically. Browser rendering runs as a dedicated non-root service on an internal-only network, uses an ephemeral sandboxed context, and can reach targets only through the validating fetch proxy. It reports CAPTCHA/challenge pages instead of trying to bypass them.

Adaptation

Infrastructure

Quick examples

Search with content extraction

curl "http://localhost:3939/search/extract?q=python+async+patterns&count=3"

Returns search results with extracted content inline — no second round-trip to /read.

Strategy search

curl "http://localhost:3939/search/strategy?q=fetch+api&mode=code&count=5"
curl "http://localhost:3939/search?q=AI+regulation&mode=academic&count=5"
curl "http://localhost:3939/search?q=Python&mode=reference&count=5"

Modes validate or call only their declared sources instead of falling back silently: general tries Bing first, then uses DuckDuckGo/Brave and direct reference/community providers only when more coverage is needed; code uses direct GitHub, MDN, Docker Hub, and PyPI providers; academic uses direct arXiv, Crossref, OpenAlex, and Semantic Scholar providers; news uses Reuters, Bing News, DuckDuckGo News, and Wikinews through SearXNG; reference uses direct Wikipedia and Wikidata providers; community uses direct Hacker News; private avoids broad general web engines.

Each strategy response includes meta.engine_attempts with source/provider, query, raw result count, latency, and upstream errors so blocked or empty providers stay visible.

Provider telemetry is available without running another probe:

curl "http://localhost:3939/providers/health"
curl "http://localhost:3939/providers/stats"

Telemetry is in-memory and reflects live attempts since the API process started. It tracks attempts, successes, empty-result rate, errors, latency, last error, and last success per direct provider or SearXNG pack.

Deep search (query expansion)

curl "http://localhost:3939/search/deep?q=ethon+industrial+ai+platform&count=10"

Server-side query variation + parallel execution + result fusion. Surfaces results that flat /search misses.

Read a URL (escalating extraction)

curl "http://localhost:3939/read?url=https://example.com/paywalled-article"

{
  "url": "https://example.com/paywalled-article",
  "content": "Full article text extracted via Wayback Machine...",
  "strategy": "wayback",
  "chars": 4821,
  "cached": false,
  "strategies_tried": ["direct", "readability", "ua_rotation", "wayback"]
}

Browser render a JS page

curl "http://localhost:3939/providers/browser/fetch?url=https://example.com/app&max_links=20"

Returns rendered text, page title, final URL, extracted links, trust metadata, and challenge_detected=true if the page is a CAPTCHA or bot challenge.

Batch read

curl -X POST "http://localhost:3939/read/batch" \
  -H "Content-Type: application/json" \
  -d '{"urls": ["https://a.com", "https://b.com", "https://c.com"]}'

Python SDK

pip install agentsearch-client

from agentsearch import AgentSearch

client = AgentSearch()  # defaults to localhost:3939
results = client.search("manufacturing OEE best practices", count=5)
for r in results.results:
    print(f"{r.title} — {r.url}")

# Content extraction
page = client.read("https://example.com/article")
print(page.content[:500])

# Browser-rendered extraction for JS-heavy target pages
rendered = client.browser_fetch("https://example.com/app", max_links=20)
print(rendered.title, rendered.links[:3])

# Batch read
pages = client.read_batch(["https://a.com", "https://b.com"])
print(f"{pages.successful}/{pages.total} succeeded")

For authenticated instances, pass token=... or use AGENT_SEARCH_TOKEN, AGENTSEARCH_TOKEN, ~/.openclaw/workspace/credentials/agent-search-token.txt, or ~/.config/agent-search/token. CWD credential files are never loaded. Remote bearer-token connections require HTTPS unless the SDK caller explicitly sets allow_insecure_http=True.

LangChain tool

from langchain.tools import tool
import requests

@tool
def web_search(query: str) -> str:
    """Search the web using AgentSearch."""
    resp = requests.get("http://localhost:3939/search", params={"q": query, "count": 5})
    return "\n".join(
        f"- {r['title']}: {r['url']}\n  {r['snippet']}"
        for r in resp.json()["results"]
    )

MCP server (Claude Desktop, Cursor, Windsurf)

pip install mcp httpx
python mcp-server/server.py

For authenticated instances, set AGENT_SEARCH_TOKEN or run:

python mcp-server/server.py --token "change-me"

Add to Claude Desktop config:

{
  "mcpServers": {
    "agent-search": {
      "command": "python",
      "args": ["/path/to/mcp-server/server.py"]
    }
  }
}

See mcp-server/README.md for details.

Private stack (Tor + encrypted DNS)

The optional private stack adds a fully anonymized search path:

┌──────────┐    ┌──────────────┐    ┌───────────────┐    ┌──────────┐
│  :3940   │───▶│ api-private  │───▶│ searxng-priv  │───▶│   Tor    │──▶ Internet
│ (agent)  │    │ (FastAPI)    │    │ (SearXNG)     │    │(Snowflake│
└──────────┘    └──────────────┘    └───────────────┘    │ + obfs4) │
                                                          └──────────┘
                 All containers use CoreDNS → Cloudflare DoT
                 tor-internal network: no direct egress possible

What this gives you:

• Your ISP sees TLS to Cloudflare (DNS) and WebRTC-looking traffic (Snowflake). Not search queries.
• The private SearXNG instance lives on an internal-only Docker network with no internet route except through Tor.
• Port 3939 = direct (fast), port 3940 = anonymized (slower, private).

Setup:

./scripts/prepare-searxng.sh
docker compose -f docker-compose.yml -f examples/compose.private.yml up -d --build

All private stack configs live in examples/ — copy and customize as needed.

Architecture

Port 3939 (direct)                    Port 3940 (Tor-anonymized)
     │                                      │
     ▼                                      ▼
┌─────────┐                          ┌─────────────┐
│   API   │                          │ api-private  │
│(FastAPI) │                          │  (FastAPI)   │
├─────────┤                          ├─────────────┤
│ dedup   │  ┌─────────────────┐     │ same code   │  ┌───────────────┐
│ scoring │  │    SearXNG      │     │ Tor egress  │  │ SearXNG-priv  │
│ cache   │──│ Google, Bing,   │     │ only        │──│ (tor-internal │
│ scrub   │  │ DDG, Brave,     │     └─────────────┘  │  network)     │
│ killchn │  │ /engines list   │                       └───────┬───────┘
│ trust   │  └─────────────────┘                               │
│ evolver │                                              ┌─────┴─────┐
└─────────┘                                              │    Tor    │
     │                                                   │ Snowflake │
     ▼                                                   │  + obfs4  │
┌─────────┐                                              └───────────┘
│ CoreDNS │──▶ Cloudflare DoT (encrypted DNS)
└─────────┘

Key modules (6,700 LOC)

Plus: 5 pluggable adapters (Cloudflare bypass, Medium, 403 handler, parse error recovery, empty content fallback), MCP server, Python SDK, test suite.

Case study: 0 → 17 frameworks per hunt

A real autonomous research agent ("the wolf") uses every AgentSearch endpoint. Before AgentSearch was wired in correctly, the agent's hand-rolled SearXNG client silently 401'd on three of four engines. Every hunt on a low-profile entity returned 0 frameworks.

After: 17 frameworks per hunt, 7/7 gaps closed. Same agent, same model, same prompts. The difference was the search infrastructure underneath.

→ Full walkthrough: case-studies/wolf.md

Configuration

Environment variables (set in docker-compose.yml or .env):

Limitations and security notes

• Search engines, news engines, rate limits, and failure modes depend on the connected SearXNG instance. /engines is the live source of truth.
• Bearer auth is a simple local API gate, not a multi-user authorization system. Treat AGENT_SEARCH_TOKEN as a shared service token.
• Rate limiting is in memory. It resets on restart and is per API process.
• Query statistics use local SQLite with WAL and a bounded lock timeout. For high-volume multi-worker deployments, move query logging to an external database or telemetry backend.
• The MCP package intentionally bounds its mcp dependency to the tested 1.27.x line. Upgrade deliberately and run the package/CI checks before publishing.
• Content extraction validates the starting URL and every redirect hop before fetching redirected content, but fetched third-party pages are still untrusted and are scrubbed before being returned.
• Google Cache is unreliable because public cache availability changes frequently.
• The Tor/private stack is intentionally slower than direct search.

Development

pip install -r requirements.txt
SEARXNG_URL=http://localhost:8080 uvicorn app.main:app --reload --port 3939

# Run tests
pytest tests/

Release and GitHub governance

• Test runs on every push and pull request.
• CodeQL runs on push, pull request, and a weekly schedule.
• Dependabot watches Python packages, Dockerfiles, and GitHub Actions.
• Version releases are created from semantic tags such as v2.0.1 or by manually running the Release workflow with a tag input.
• Update CHANGELOG.md before creating a release tag.

Contributing

1. Fork → branch → commit → PR.

Issues and PRs welcome. If you're building an agent that needs search, this is for you.

License

The root AgentSearch API, SDK, Docker stack, and docs are MIT licensed. The MCP server under mcp-server/ is AGPL-3.0 licensed; see mcp-server/LICENSE.