reapfield
Extracts structured JSON from any URL using plain-language field specs, with caching and LLM-based selector discovery. Exposes scrape, list_cached_selectors, and refresh_selectors tools over stdio, and blocks private addresses for safety.
README
reapfield
(No PyPI badge yet — nothing is published there. See Install.)
Give it a URL and a plain-language field spec, get structured JSON back — including on pages that only render under JavaScript.
$ git clone https://github.com/PedroHenriqueNS/reapfield.git && cd reapfield
$ uv sync
$ export ANTHROPIC_API_KEY=sk-...
$ uv run reapfield https://books.toscrape.com/catalogue/a-light-in-the-attic_1000/index.html \
--fields "title, price:float, in_stock:bool"
reapfield: mode=one records=1 llm_calls=1
{
"title": "A Light in the Attic",
"price": 51.77,
"in_stock": true
}
That first run cost one LLM call — this page has no JSON-LD or other structured data, so
reapfield had to derive selectors. Run the identical command again and it's free:
$ uv run reapfield https://books.toscrape.com/catalogue/a-light-in-the-attic_1000/index.html \
--fields "title, price:float, in_stock:bool"
reapfield: mode=one records=1 llm_calls=0
{
"title": "A Light in the Attic",
"price": 51.77,
"in_stock": true
}
The idea
Every one-off scraper gets written twice: once to find the selectors by hand, and again when the site changes.
reapfield pays an LLM once per field per domain to discover the CSS selectors, caches
them, and replays them deterministically forever after. It goes back to the LLM only when a
cached selector actually stops working. Steady-state cost is zero LLM calls — that
llm_calls=1 → llm_calls=0 transition above is the whole design in one line.
Before it ever considers an LLM it tries the free paths in order: JSON-LD, OpenGraph and
<meta>, __NEXT_DATA__, __NUXT__, inline JSON, then config-pinned selectors, then the
cache. Plenty of pages never reach the paid step at all.
Not a goal: defeating anti-bot systems. When a site refuses automated access,
reapfield names the system and stops.
Install
From source — this is the path that actually works today:
git clone https://github.com/PedroHenriqueNS/reapfield.git
cd reapfield
uv sync
uv run playwright install chromium # only needed for JS-rendered pages
export ANTHROPIC_API_KEY=sk-... # only needed the first time a domain is scraped
Everything below assumes you're in that cloned directory, running commands with uv run.
Not yet on PyPI. pip install reapfield, uv tool install reapfield and
uvx reapfield are the intended install path once a release ships, but nothing is
published there yet — the commands above are correct on first release and dead until then.
Usage
reapfield <url> --fields "title, price:float, in_stock:bool"
[--format json|jsonl|csv] [--one | --many]
[--strict] [--refresh] [--no-llm] [--max-llm-calls N]
[--no-cache] [--cache-ttl SECONDS]
[--scroll N] [--paginate N] [--allow-private]
Types are optional and inline: price:float, in_stock:bool, count:int. Unannotated
fields are strings. Field names are arbitrary — --fields "reactor_id, coolant_temp_c:float"
takes exactly the same code path as title, price.
Exit codes: 0 at least one field extracted · 1 nothing extracted, or --strict with any
miss, or blocked, or robots-disallowed, or an --one/--many conflict · 2 usage error.
A field that could not be found comes back as null, with the reason on stderr. That is a
result, not a crash.
How invalidation works
There is no TTL on a selector. A selector that still works is still correct, and expiring it just buys LLM calls.
A cached selector is dropped and re-derived when it matches nothing, or when the value
it returns fails the declared type — price:float suddenly yielding "Add to basket"
means the page moved under us. That second case is the interesting one: nothing 404'd and
nothing expired, the selector simply started pointing at the wrong node.
MCP server
Same core as the CLI, exposed over stdio. Run this from the cloned project directory so
uv run resolves to this checkout:
claude mcp add reapfield --scope local -- uv run reapfield-mcp
Tools: scrape, list_cached_selectors, refresh_selectors, prepare_issue_report. The
last one drafts a bug report about reapfield itself and returns a link — it never submits;
a human opens the link.
CLI URLs come from you; MCP URLs come from a model that may be acting on text it read off a
web page. That is why private, loopback and link-local addresses are blocked by default
on the CLI too, not only over MCP — including on redirects. A redirect target is chosen
by the server you were pointed at, not by whoever typed the original URL, so it earns no
more trust: redirects are followed by hand, capped at 5 hops, with this check, robots.txt,
rate limiting and the gated-platform guard re-applied at every hop. 169.254.169.254,
the cloud metadata endpoint, is the case that matters most. Checking the resolved IP
rather than the hostname string is what stops DNS rebinding.
Opt out with --allow-private on the CLI (for scraping localhost or a LAN host during
development) or REAPFIELD_MCP_ALLOW_PRIVATE=1 for MCP.
Config
~/.config/reapfield/config.toml, then ./reapfield.toml overriding it per key.
concurrency = 4
user_agent = "reapfield/0.1 (+https://github.com/PedroHenriqueNS/reapfield)"
[domains."books.toscrape.com"]
fetcher = "http" # auto | http | browser
rate_limit = 2.0 # req/sec; robots.txt Crawl-delay is still a floor
wait_for = ".product_main" # browser mode only
pagination = "li.next > a"
[domains."books.toscrape.com".selectors]
price = ".price_color" # pinned: never derived, never touched by --refresh
_row = "article.product_pod" # the repeating container, for --many
[contribute]
reports = "ask" # ask | never -- see below
Pinned selectors live in a separate store from the cache, so --refresh can never
overwrite a decision you made by hand. Append @attribute to read one: "h3 a@title".
The MCP server can invite a connected agent to draft a bug report about your session (see
prepare_issue_report above). [contribute] reports = "never", or the equivalent
REAPFIELD_ISSUE_REPORTS=never env var, turns that off entirely; the env var wins if both
are set. Default is "ask".
Manners
robots.txt is obeyed, with no override flag. Every domain is rate limited, and a
Crawl-delay in robots.txt always wins over the config. The user agent is honest and
identifiable, never randomized.
Development
uv run pytest # offline: no network, no API key
The suite stubs the derivation function, so nothing in it can reach Anthropic or the
network. See SPEC.md for the full design and docs/adapters.md for the adapter seam.
Contributing
See CONTRIBUTING.md. Security issues go through SECURITY.md — please do not open public issues for them. Release history is in CHANGELOG.md.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.