A self-healing web scraper. The LLM writes the crawler once — then your pages run free.

Stop paying an LLM on every page, and stop scrapers that break silently when a site redesigns. crawloop compiles a cheap deterministic crawler, serves instantly via the LLM the moment one breaks, and regenerates a fresh version in the background — back to $0.

Working proof of concept — tested live on a real website: the loop learns a free deterministic crawler from books.toscrape.com over real HTTP, then extracts a fresh page for $0 (no model call). Backed by 546 offline tests; provider-agnostic via litellm (OpenAI / Anthropic / Gemini, auto-detected); falls back to the LLM when it can't compile a family, never worse.

• $0 and milliseconds per page in steady state — the model is a compiler, not a runtime. It runs once to write the crawler, never to serve a request.
• A redesign is never an outage and never silent — drift is detected, the page is served now via the LLM, and a new crawler version is promoted automatically.
• Tested on a real site, not a toy — the loop learns a free crawler from live books.toscrape.com and extracts fresh pages for $0; the whole cycle also reproduces offline in ~30s so you can watch the mechanism end-to-end.

# Run the loop against a REAL site, then serve fresh pages free (needs a provider key):
RUN_LIVE_LLM=1 python -m pytest tests/test_live_llm_smoke.py -q   # live end-to-end on books.toscrape.com

# Or watch the full break → serve → heal → free cycle reproduced offline (no key, no network):
python examples/selfheal_demo.py

How it works → · Quickstart → · See it heal → · The tradeoff →

---

Why this exists

LLM-per-page scrapers are seductive — point a model at HTML, get JSON. But in production they have two structural problems that never go away:

• You pay per page, forever. Every page, every re-crawl, every run hits the model. At a few cents a page that's real money at scale — and unlike code, the bill never amortizes. Crawl a million pages and you pay a million times.
• They break silently. When a site redesigns, an LLM doesn't know it broke. It confidently extracts the wrong thing (or nothing) and raises no error — there is no drift signal, so you find out from downstream garbage, days later.

crawloop flips the model. The LLM is a compiler and a teacher, not a runtime. It writes a deterministic crawler once, acts as the oracle that grades regenerated versions, and steps in only during a breakage to serve data while a fresh crawler is built. Steady state runs on free, instant, byte-reproducible code.

The contract, in five words: serve now, heal in the background.

---

Quickstart

git clone https://github.com/Jimmynycu/Crawloop.git
cd Crawloop
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

The real-world test. Point the loop at a live website; it samples real pages, the LLM compiles a deterministic crawler, the gauntlet promotes it, and the promoted crawler then extracts a fresh real page for $0 (no model call):

RUN_LIVE_LLM=1 OPENAI_API_KEY=sk-...  python -m pytest tests/test_live_llm_smoke.py -q
# end-to-end on live books.toscrape.com — real HTTP + a real model (a few cents)

Reproduce the mechanism offline (no key, no network) — the same engine, with a committed cassette standing in for the model, so the whole cycle is deterministic and free to watch:

python examples/selfheal_demo.py

That narrated demo — and the matching end-to-end test (tests/test_selfheal_e2e.py) — drives the real engine through:

Step	What happens	Cost
1 · Fast path	a healthy generated crawler extracts records with no LLM call	$0
2 · Break	the fixture site's layout is mutated (a simulated redesign)	—
3 · Serve now	drift is detected; the page is served immediately via the T2 LLM fallback against the schema	paid, once
4 · Heal	the Loop samples pages, uses the LLM as an oracle, gauntlet-scores candidates, and promotes a v2	paid, once
5 · Reuse	the next request runs the healed crawler	$0 again
6 · Recover	a 403 block is hit, the per-domain access ladder escalates, gets through, and saves the winning strategy	—

python -m pytest   # the full suite — 546 tests, all offline, all without a key

Note

Real runs (against your own authorized sites) need an API key for the T2 fallback and the Loop. The demo above proves the machinery first, for free.

---

How it works

A request flows through a version ladder of cheap deterministic crawlers first; the LLM is only ever reached on a real breakage.

Authorize (allowlist gate) → route to a registered page family → run that family's version ladder of generated crawlers (the cheap, fast path). If a version validates, items are served with no LLM call.

If every version fails, the failure is classified:

• Drift → served now by T2 (the LLM reading the HTML against the schema) while the regeneration Loop rebuilds a crawler in the background.
• Block (429 / login wall / anti-bot) → the access-recovery ladder escalates and retries.
• Transient error → retried.
• Gone (404/410) → stops.

See the full runtime architecture → docs/design.html#arch — tiers T0/T1/T2/Loop/Access, the two self-healing loops, and the safety model.

---

See it heal

The offline demo, as it actually runs — a redesign breaks the crawler, the request is served anyway, a new version is written and promoted, and steady-state model calls go back to zero. The model-calls column is the whole story: 0 → 6 (paid once) → 0.

Run it yourself in ~30s, no API key: python examples/selfheal_demo.py — the real engine drives the full cycle (a committed cassette stands in for the LLM). The numbers above are its actual output.

---

The design tradeoff

The table contrasts the two architectures — not a benchmark, no measured numbers from any system. It is the structural argument for compiling a crawler instead of calling an LLM on every page; the axes follow directly from "code runs vs a model runs" and from "the system has a drift signal vs it doesn't."

Dimension	crawloop	Generic LLM-per-page	Why
Cost model	compile once, then run free	a model call on every page, forever	the LLM bill amortizes for crawloop, never for per-page
Latency	code (parsel) — local, no round-trip	an LLM round-trip per page	deterministic code has no network step in steady state
Determinism	byte-identical output for the same page	may vary run-to-run	code is deterministic; sampling is not
Drift handling	detects validation drift → self-heals	no signal; ships wrong data blind	crawloop validates each extraction and knows when it broke
Worst case	safely falls back to the LLM = parity	—	a family it can't compile is served by the LLM, never worse

Honest counterpoint: the oracle now reads wide, deeply-nested JSON islands in full (proven live), and the loop escalates the model to clear the bar — but compiling a very wide, normalized schema to fully deterministic code is still the hard part. When crawloop can't compile a family to the bar it falls back to the LLM (= parity), spending only the one-time bootstrap.

---

Features

• Compile-once extraction — LLM-generated deterministic Python crawlers; steady state runs at $0 and milliseconds per page.
• Self-healing on drift — a layout change triggers an instant LLM fallback plus background regeneration of a new crawler version. No outage, no silence.
• Version ladder, not overwrite — each family keeps an ordered v1, v2, v3… of immutable crawlers; healing appends a version and flips the active pointer (handles gradual redesigns and A/B layouts), with one-command rollback.
• Access recovery — a 429 / login wall / anti-bot block isn't terminal: an ordered, per-domain ladder (backoff → stealth browser → session → bypass token) escalates until one gets through, and the winning strategy is saved and reused.
• Hard allowlist, enforced on every hop — no URL outside authorized_domains.yaml can ever be fetched; cross-host/SSRF redirects are refused.
• Sandboxed generated code — every candidate crawler is AST-checked (import/call allowlist, no dunder escapes) and run in a resource-capped subprocess before it can touch a real page.
• Pluggable Pydantic schemas — drop a BaseModel in schemas/; it's auto-registered as Name@1. Mark VOLATILE fields so the validator compares price/stock tolerantly.
• Full audit trail — every promotion and access recovery is recorded (SQLite + audit.jsonl): what the system did, and why, reviewable after the fact.
• Provider-agnostic — model calls go through litellm; use OpenAI, Anthropic, or Gemini, auto-detected from whichever API key is set (or pin one with --model / CRAWLOOP_MODEL).
• Tiered model escalation — the loop runs a cheap model by default and escalates the regeneration (oracle + codegen) to a stronger model only when no candidate clears the gauntlet; the promoted crawler is still free deterministic code, so the one-time cost amortizes. (See the note below.)
• Get the data out — crawloop crawl --json returns the extracted records (not just a count); from crawloop import Engine, run_loop embeds the engine as a library.

---

Loop engineering: tiered model escalation

A small technique that lets the cheap default actually work. The regeneration loop is graded by a strict gauntlet — a candidate crawler is promoted only if it agrees with the LLM oracle on every item (≥0.98 per-item, never a mean), the record counts match exactly, and it holds across ≥3 independent samples. A cheap model is plenty for the oracle and the fast path, but its codegen may not clear that bar on a hard multi-record page.

So the loop escalates: it regenerates with the cheap model first, and only if nothing clears the gauntlet does it retry the whole regeneration (oracle + codegen) once with a stronger model (crawloop/llm.py escalation_model, wired in crawloop/loop/driver.py). Whatever wins is still free, deterministic code, so the one-time stronger-model cost amortizes over every future page.

In plain terms: spend the smart model once, only when the cheap one can't prove it's right — then run for free forever. Proven live: starting from gpt-4o-mini, the loop auto-escalates to gpt-4o and promotes a free, multi-record crawler.

---

Install

Requires Python 3.12+.

python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
python -m pytest  # 546 tests, no API key needed

The access ladder's browser rungs use a real PlaywrightBrowserRunner / StealthBrowserRunner (crawloop/browser.py) that re-enforces the allowlist on every navigation and in-page redirect (the browser bypasses the guarded HTTP client, so it gates itself). Install the browser binaries once with playwright install. Gated live browser tests live in tests/test_browser_live.py (RUN_BROWSER_TESTS=1).

Environment variables. No secret is ever stored in the repo or config; the config only names the env var to read.

• A provider API key — any one of OPENAI_API_KEY, ANTHROPIC_API_KEY, or GEMINI_API_KEY enables real runs (the T2 fallback and the Loop, via litellm). crawloop auto-detects which is set and picks a sane default model; override with --model (e.g. --model openai/gpt-4o) or CRAWLOOP_MODEL. Not needed for the test suite or for --offline on a healthy family.
• Per-domain credentials / tokens — named by the *_env fields in access_strategies (e.g. session → creds_env, bypass_token → value_env, proxy_env), read from the environment at fetch time.

---

CLI usage

Installed as crawloop (entry point crawloop.cli:main). Global options (--config, --db, --crawlers-dir, --fixtures-dir) default to authorized_domains.yaml and a local .crawloop/ working dir.

# Crawl one URL through the full engine (authorize → route → ladder → heal):
crawloop crawl https://books.toscrape.com/catalogue/page-1.html --schema Product@1

# Same, machine-readable:
crawloop crawl https://books.toscrape.com/catalogue/page-1.html --schema Product@1 --json

# Inspect the registry:
crawloop family list
crawloop family show books.toscrape.com/product_list

# Run the regeneration loop by hand (seeds = pages to sample):
crawloop loop run books.toscrape.com/product_list \
    https://books.toscrape.com/catalogue/page-1.html \
    https://books.toscrape.com/catalogue/page-2.html

# Inspect the per-domain access store and the audit trail:
crawloop access status
crawloop audit                                    # all events
crawloop audit books.toscrape.com/product_list    # one family

# Add --offline to forbid constructing a real model/browser (a healthy family's
# fast path needs neither; a drift/bootstrap then fails loudly, not over the network):
crawloop crawl https://books.toscrape.com/catalogue/page-1.html --offline

--schema is required only for a new (unrouted) URL the engine has to bootstrap from; a known family uses its stored schema.

Adding a schema

Output schemas are plain Pydantic models contributed as .py files in schemas/. Drop in a BaseModel subclass; it's auto-discovered and registered under f"{ClassName}@1". Use extra="forbid" so unexpected keys are caught, and declare VOLATILE for fields that change often so the validator compares them tolerantly:

# schemas/product.py
from typing import ClassVar
from decimal import Decimal
from pydantic import BaseModel, ConfigDict, Field, HttpUrl


class Product(BaseModel):
    model_config = ConfigDict(extra="forbid")
    name: str = Field(min_length=1)
    price: Decimal = Field(gt=0, lt=1_000_000)
    currency: str = Field(pattern=r"^[A-Z]{3}$", default="GBP")
    in_stock: bool
    url: HttpUrl
    image_url: HttpUrl | None = None
    VOLATILE: ClassVar[set[str]] = {"price", "in_stock"}

authorized_domains.yaml — the hard allowlist + crawl policy

This file is the operator's explicit, mandatory allowlist. Every fetch must pass AppConfig.assert_authorized, so a URL whose host isn't listed can never be requested. It also carries per-domain policy: rate limit, JS rendering, and the ordered access strategies tried when a block is hit.

respect_robots: false          # POC default (see Authorized use). Flip to honor robots.txt.

domains:
  - domain: books.toscrape.com
    max_rps: 1.0
    render_js: false

  - domain: shop.example.com
    max_rps: 0.5
    render_js: true
    note: "owned by us; authorized 2026-06-12"
    access_strategies:         # ordered ladder; recovery walks it and persists the winner
      - backoff                #   plain GET with exponential backoff on 429/5xx
      - stealth_browser        #   patched headless browser render
      - session: { login_url: "https://shop.example.com/login", creds_env: "EXAMPLE_LOGIN" }
      - bypass_token: { header: "x-waf-bypass", value_env: "EXAMPLE_WAF_TOKEN" }
    proxy_env: "EXAMPLE_PROXY_URL"

---

Authorized use

This is a tool for crawling sites you own or are explicitly authorized to crawl. It is deliberately not a general-purpose scraper for sites you have no relationship with.

• The allowlist is mandatory. authorized_domains.yaml is a hard gate on every fetch (including every redirect hop — a cross-host/SSRF redirect to an unlisted host is refused). No override.
• respect_robots defaults OFF because the intended targets are owned/authorized properties. Flip it to true to honor robots.txt; decide deliberately per deployment. (Note: the flag is parsed but not yet enforced — see the roadmap.)
• The CAPTCHA rung is opt-in and authorized-domains-only. The system never auto-defeats a captcha: captcha_solver raises unless an operator has explicitly set authorized: true for that domain and wired a provider (none ships here). Stealth browser and bypass tokens are likewise explicit per-domain opt-ins — courtesy headers and rate limits are the default, not evasion.

If you would not be comfortable explaining a crawl to the site's owner, it does not belong on the allowlist.

---

Roadmap

Stated candidly — these are the gaps between "promising POC" and "drop-in replacement."

Recently landed — each in plain English, with where it's proven:

• The oracle now reads wide JSON islands in full. A flat 8000-char prompt cap used to chop the gold record off a 100K+ __NEXT_DATA__ island, so the loop could never promote on those pages. The hoisted JSON now gets its own budget (crawloop/htmlutil.py json_max_chars, crawloop/fallback.py CRAWLER_LOOP_ORACLE_JSON_CAP). Proven live (a record buried 53K deep was extracted by a real model).
• The oracle absolutizes URLs. It now receives the page URL and resolves relative hrefs to absolute, so listings whose url field was failing validation now pass. Proven live.
• Tiered model escalation (the new loop technique) — cheap model by default; escalate the regeneration to a stronger model only when the gauntlet isn't cleared. Proven live (gpt-4o-mini → gpt-4o → promote). See the section above.
• Provider-agnostic — OpenAI / Anthropic / Gemini, auto-detected; --model / CRAWLOOP_MODEL override (was hardwired to Anthropic).
• Get the data out — crawloop crawl --json returns the records; from crawloop import Engine, run_loop is a real library API.
• Live-model smoke tests — env-gated (RUN_LIVE_LLM=1, tests/test_live_llm_smoke.py); the loop is no longer offline-only.
• Plus: the core-deterministic + LLM-tail hybrid (crawloop/hybrid.py), the real BrowserRunner (Playwright + Patchright), and clean wheel packaging.

Next up — the honest deltas vs the mature tools (Firecrawl / Kadoa / Crawl4AI):

• Prompt → schema — describe fields in prose and have the model write the schema (today you drop a Pydantic file in schemas/).
• Deterministic multi-record discovery — a free, no-LLM path for JSON-array listings (today listings promote via codegen + escalation, which works, but isn't the zero-LLM path).
• Enforce respect_robots — still parsed, not enforced (the authorized-only ethos lessens it, but it should be honored when on).
• Managed proxies, scale, dashboards, SDKs, PyPI publish — the operational surface the funded tools have; out of scope for a one-author POC, and called out so the comparison stays honest.

Intentionally out of scope for this POC (Phase 2): non-LLM fingerprint healing (T1), DOM-shingle family routing, sampled production LLM-judge, distribution monitors + scheduled canaries, a web dashboard, Postgres, and concurrency hardening.

---

Contributing

PRs welcome — especially the open roadmap items above (oracle reliability on huge JSON islands and JSON-first codegen are the highest-impact right now).

1. Fork, branch, and pip install -e ".[dev]".
2. Run python -m pytest (546 tests, no key needed) and ruff check . — both must stay green.
3. Add tests for your change; the offline fixture server (tests/fixture_server/) lets you exercise the full loop deterministically.
4. Open a PR describing the behavior change and how you verified it.

Found a bug or have a design question? Open an issue. See CONTRIBUTING.md for the full guide.

---

License

Released under the Apache License 2.0 — see LICENSE.

_{Built to demonstrate the self-heal + access-recovery loop on sites you own or are authorized to crawl. If it saved you from an LLM bill that never ends, a ⭐ helps others find it.}