wesearch🕸️

A wascally web research toolkit for agents; web search, web fetch, and publication research.

Quick Start

# Mac:
#   # Required for quick install.
#   brew install uv

# Ubuntu/Debian:
#   # Required for quick install.
#   sudo apt-get install -y curl
#   curl -LsSf https://astral.sh/uv/install.sh | sh

uv add wesearch

# Alternatively: python -m pip install wesearch

wesearch is a synchronous, batteries-included library for programmatic web access: run a search, fetch a page through a real-browser fingerprint, extract content, and look up scholarly papers across multiple providers. It is the web layer factored out of a larger agent stack, so it is built to survive bot-detection, rate limits, and flaky endpoints without a running browser in the common case.

Example

from wesearch.search.search import search
from wesearch.fetch import RequestParams, fetch
from wesearch.scrape import get_element_content

# Web search (DuckDuckGo by default)
hits = search("denoising recursion models", num_results=10)
for r in hits:
    print(r.title, r.url)

# Fetch + scrape
body, _session = fetch("https://example.com", request=RequestParams(timeout_sec=10))
title = get_element_content(body.decode("utf-8"), "h1")

# Scholarly papers: Semantic Scholar + OpenAlex, reciprocal-rank-fused by default
from wesearch.paper.search import search as paper_search
from wesearch.paper.ids import normalize_id
from wesearch.paper.details import metadata

result = paper_search("attention is all you need", limit=5)
for rec in result.records:
    print(rec.title, rec.year)

meta = metadata(*normalize_id("arXiv:1706.03762"))

Each name is imported from the submodule that defines it; the top-level __init__ re-exports nothing.

What's inside

wesearch/
├── types/           the vocabulary every layer shares; imports nothing internal
│   ├── params.py    RequestParams + Content/Retry/Observe/Policy; Transport,
│   │                  Extractor, Trust
│   ├── extractor.py the Extract protocol each extractor satisfies
│   └── errors.py    FetchError, BotDetectionError + subclasses
├── fetch/           the sole HTTP egress
│   ├── fetch.py     fetch(url, request=RequestParams(...)) -> (body, session)
│   ├── transport/   how bytes are retrieved
│   │   ├── curl.py      curl-cffi transport (TLS/JA3 browser impersonation)
│   │   ├── stdlib.py    dependency-free urllib transport
│   │   ├── zendriver.py opt-in real-Chrome backend for JS-gated pages
│   │   └── transport_routing.py  per-domain transport selection
│   ├── extractor/   how a fetched page becomes text
│   │   ├── html2text.py   every text node as Markdown
│   │   ├── markdownify.py the document's elements as Markdown
│   │   ├── trafilatura.py the scored article body only (the default)
│   │   └── raw.py         the source, untouched
│   ├── providers/   per-site fetch strategies (reddit, google_news, x, ...)
│   └── challenge.py bot-challenge detection and classification
├── search/          web search over pluggable backends
│   └── search.py    search(...) over SearXNG / DuckDuckGo / Google;
│                      SearchResult / PaperResult / ImageResult
├── paper/           scholarly-paper lookup
│   ├── search.py    search(...) across Semantic Scholar, OpenAlex, SearXNG
│   ├── details.py   metadata / references / citations
│   ├── authors.py   author search and publication lists
│   ├── fetch.py     PDF download cascade
│   └── providers/   per-source backends (openalex, s2, searxng)
├── mcp/             the MCP surface; the only place the mcp SDK is imported
│   └── server.py    wesearch-mcp, one tool per public function
├── chrome/          real-browser fingerprints
│   ├── headers.py   Chrome request headers (incl. x-browser-validation)
│   └── useragents.py  vendored, refreshable User-Agent pools
├── web.py           fetch_web(...): fetch + provider dispatch + extraction
├── profile.py       cross-process per-(egress_ip, domain) cookie + UA jar
├── ratelimit.py     cross-process, per-domain rate limiting
└── scrape.py        get_element_content(html, selector)

Transports

fetch picks a transport per domain:

• curl-cffi (default): TLS/JA3 browser impersonation, no browser process.
• stdlib: dependency-free urllib fallback.
• zendriver: an opt-in headless-Chrome backend for JavaScript-gated pages, used only for domains that require it.

A persistent per-(egress_ip, domain) profile (cookies + User-Agent) is loaded and saved transparently, and cross-process rate limiting paces requests so concurrent workers stay under each site's threshold.

API keys & configuration

Everything works keyless out of the box; the environment variables below raise your rate limits or unlock a backend, and are all optional unless noted.

• SEMANTIC_SCHOLAR_API_KEY -- optional. Without it, Semantic Scholar lookups (paper.search, paper.details) share a low-rate public tier, and paper.search(..., source="fused") may return complete=False when S2 throttles the request. Set it for a higher-throughput tier.
• OPENALEX_EMAIL -- optional. Identifies you to OpenAlex's "polite pool" for more headroom than anonymous requests.
• OPENALEX_API_KEY -- optional. A higher OpenAlex request budget.
• SEARXNG_URL -- required only to use a "searxng" backend (search(..., backend="searxng") or paper.search(..., source="searxng")); the base URL of a SearXNG instance you control or trust.

State (the cookie/User-Agent profile jar, cross-process rate-limit lockfiles, the browser transport's persistent Chrome profile) is written under the OS's standard per-user data directory -- XDG_DATA_HOME (or ~/.local/share) on Linux, ~/Library/Application Support on macOS, %LOCALAPPDATA% on Windows -- namespaced per component (see wesearch/lib/userdirs.py). No configuration file is required or read.

The "zendriver" transport needs a non-snap Chrome or Chromium (e.g. Google Chrome's .deb on x86_64). Snap-packaged Chromium fails every launch with BrowserUnavailableError for two reasons, neither of which the library can work around:

1. The snap wrapper takes several seconds to expose DevTools -- far beyond the launch budget (0.5s per attempt, 6 attempts), which is sized so the curl-then-zendriver cascade fails fast on hosts with no usable browser rather than stalling every fetch.
2. Snap's AppArmor confinement silently blocks writes under hidden home paths like ~/.local/share, so Chrome dies on its profile lock wherever the profile jar lands by default.

Development

uv sync --all-groups

Tests are tiered with pytest markers; the default run (uv run pytest) executes only the fast unit tier:

addopts = -m 'not ci_smoke and not cuda and not integration and not performance and not cluster and not slow'

• ci_smoke -- slower package smoke tests, run explicitly in CI.
• cuda -- requires a real CUDA device.
• integration -- requires networking or external CLIs.
• performance -- timing-sensitive.
• cluster -- requires live cluster access.
• slow -- expensive local correctness tests (JIT, full fixtures, git, bash, a fresh interpreter).
• real_llm -- spawns a live LLM CLI; skipped unless RUN_REAL_LLM=1.

Run a specific tier with uv run pytest -m integration, or everything with uv run pytest -m ''.

The zendriver backend tests need a Chrome or Chromium binary on PATH; without one they are skipped automatically. No other system dependency is required to run the fast unit tier.

MCP server

The toolkit is directly callable by coding agents over MCP:

pip install 'wesearch[mcp]'
wesearch-mcp                  # serves stdio; register it with your MCP client

For Claude Code: claude mcp add --scope user wesearch -- wesearch-mcp.

Tools exposed: paper_search (fused Semantic Scholar + OpenAlex), paper_details, paper_references, paper_citations, paper_pdf (downloads into the user cache and returns the path), author_search, author_papers, web_search, and web_fetch (extracted page text, with an opt-in headless-browser fallback for bot-walled sites). Outputs are deliberately compact for model consumption; abstracts are truncated and empty fields dropped. The server is synchronous and per-client — state that must be shared (rate limits, cookie/UA profiles) is already cross-process safe on disk, so no daemon is needed.

See also

Sibling projects in the rekursiv-ai family:

• sagent — The self-mutating multi-provider coding-agent CLI and typed Python library.
• trackinizer — Centralized agent database for tracking inquiries, work, and the evidence behind conclusions.
• madcatter — Rich-based Markdown renderer for the terminal; ships the mdcat CLI.
• priml — Composable PyTorch building blocks: models, optimizers, losses, and a step-based training loop.
• configgle — Hierarchical experiment configuration in typed pure-Python dataclasses instead of YAML.
• copybarista — Bidirectional source sync for publishing OSS-ready trees from a monorepo.
• sudoku — Sudoku-Extreme solved end to end with a 7M-parameter recursive transformer.

Citing

If you find our work useful, please consider citing:

@misc{rekursivai2026wesearch,
      title={Wesearch - A wascally web research toolkit for agents.},
      author={Joshua V. Dillon},
      year={2026},
      howpublished={Github},
      url={https://github.com/rekursiv-ai/wesearch},
}