A Python framework for building structured, resumable web crawlers — designed for domains where data quality matters.

What is Ladon?

Ladon enforces typed domain objects at every stage of the crawl pipeline through the SES protocol (Source / Expander / Sink). The difference from Scrapy — a proven, mature tool — is structural: instead of weakly typed scrapy.Item fields, you define typed dataclasses at the protocol level (e.g. a CommentRecord with enforced field types). The output is structured and typed without a post-processing step. This matters when the destination is an LLM training pipeline or any domain where schema correctness is not optional.

The built-in HTTP layer handles retries, exponential back-off with optional full-jitter, 429/503 Retry-After respect, per-domain rate limiting, circuit breaking, static and rotating proxy support, and robots.txt enforcement — so adapter authors focus on domain logic, not infrastructure.

Quick start

The canonical example is ladon-hackernews — an adapter that crawls the HN top-stories list and writes comments to DuckDB:

pip install ladon-crawl ladon-hackernews
ladon-hackernews --top 30 --out hn.db

No authentication. No external server. 30 stories and their comments in under a minute.

The LLM training pipeline

ladon-hackernews --top 500 --out hn.db
    → export_parquet("hn.db", "hn.parquet")
        → training pipeline

HN comments are structured, human-authored, and high signal-to-noise. The full pipeline from install to Parquet takes under five minutes. Each run writes a ladon_runs audit table to the DuckDB file — re-running skips stories already marked done, giving you resumable crawls for free.

from ladon_hackernews import export_parquet
export_parquet("hn.db", "hn.parquet")

Writing your own adapter

ladon-hackernews is the canonical reference for building an adapter. Adapters implement the SES protocol structurally — no inheritance from any Ladon base class is required. The three components to implement are:

• Source — discovers the list of root references to crawl
• Expander — maps a reference to a domain record and child references
• Sink — receives each leaf record for persistence or downstream use

See the adapter authoring guide and ADR-003 for the full protocol specification. The ladon-hackernews source is the worked example.

CLI reference

ladon info
ladon run --plugin MODULE:CLASS --ref URL [--respect-robots-txt]
ladon --version

ladon run flags:

Exit codes: 0 success · 1 fatal error · 2 partial failures · 3 data not ready (retry later)

ladon run uses default HttpClientConfig settings. For retries, rate limiting, circuit breaking, or a persistence layer, call run_crawl() directly from Python — see ladon-hackernews — Use as a library for a full example.

Async crawling

async_run_crawl() is the asyncio-native counterpart to run_crawl(). Phase 1 (expander traversal) runs sequentially; Phase 3 issues leaf fetches concurrently behind asyncio.Semaphore(config.async_concurrency) (default 10):

import asyncio
from ladon import AsyncHttpClient, async_run_crawl
from ladon.networking.config import HttpClientConfig
from ladon.runner import RunConfig

async def main() -> None:
    config = HttpClientConfig(retries=2, timeout_seconds=10)
    async with AsyncHttpClient(config) as client:
        result = await async_run_crawl(
            top_ref=my_ref,
            plugin=my_async_plugin,
            client=client,
            config=RunConfig(async_concurrency=20),
            on_leaf=my_async_persist,
        )
        print(f"consumed {result.leaves_consumed}, failed {result.leaves_failed}")

asyncio.run(main())

AsyncHttpClient mirrors all policies of HttpClient (retries, backoff, Retry-After, circuit breaker, proxy rotation, auth) using httpx as the backend. Adapters implement AsyncCrawlPlugin, AsyncSource, AsyncExpander, and AsyncSink — the same structural-protocol pattern as the sync stack.

Status

v0.2.0 — async crawling milestone. AsyncHttpClient, AsyncCrawlPlugin, and async_run_crawl() are stable and fully tested. The sync API is unchanged.

What is in v0.2.0:

• Async crawling — async_run_crawl() + AsyncHttpClient (httpx backend)
• Async plugin protocols — AsyncSource, AsyncExpander, AsyncSink, AsyncCrawlPlugin
• RunConfig.async_concurrency — bounded leaf-fetch concurrency (default 10)

What was in v0.1.0:

• HTTP 429/503 Retry-After respect and full-jitter backoff
• Static and rotating proxy support (ProxyPool, RoundRobinProxyPool)
• HTTP authentication — Basic, Digest, any requests.auth.AuthBase
• Default query parameters via HttpClientConfig(default_params=...)

What was in v0.0.1:

• SES protocol (Source / Expander / Sink) with structural typing
• run_crawl() runner with leaf isolation and RunResult summary
• HttpClient with retries, back-off, rate limiting, circuit breaker, robots.txt
• Storage protocol with LocalFileStorage
• Repository and RunAudit persistence protocols with NullRepository
• ladon run / ladon info CLI

What is coming:

• ladon-mimir — async Wikipedia adapter for LLM fine-tuning (issue #96)
• Async robots.txt enforcement in AsyncHttpClient
• Structured logging baseline (ADR-009)

Contributing

The plugin protocol is settled — contributions are welcome. Please read the documentation for design context (ADRs, plugin authoring guide) before sending a pull request.

A CLA signature is required for external contributors. The bot will prompt you on your first PR.

License

Ladon is released under the GNU Affero General Public License v3.0 only (AGPL-3.0-only). See LICENSE for the full text.

AGPL was chosen to ensure that improvements to the core framework — including when deployed as a networked service — remain open and available to the community. A commercial licence is available for organisations that cannot accept the AGPL terms — see LICENSE-COMMERCIAL.

ladon-hackernews is separately licensed under Apache-2.0.