CrawlKit Platform Model
CrawlKit is mission control for hard-to-get data: acquisition workbenches on top of a trustworthy data engineering backbone. It turns live web crawls, deterministic spiders, scrape jobs, ETL, social signals, synthetic generation, and domain applications into governed, queryable, lineage-backed data assets.
The operating loop: Acquire → Extract → Refine → Serve → Prove
Acquire
Live web crawls, spiders, scrape jobs, frontier queues, stealth/proxies,
WARC/Common Crawl, file/API ETL, social monitors, OpenRate MRF/TiC ingestion.
Extract
Declarative schemas, selectors, vertical templates, content classification,
parser stages, field processors, and deterministic structured records.
Refine
Enrichment, taxonomy, scoring, deduplication, decontamination, quality gates,
validation, outlier detection, and synthetic verification.
Serve
Datasets, query views, DataFusion SQL, DuckDB hybrid SQL, Arrow Flight,
streams, exports, reports, training formats, and API applications.
Prove
Durable workflow runs, lineage, policy, consent, audit, telemetry,
evidence drawers, governance approvals, and measurement feedback loops.
The backbone
The backbone is the persistent trust layer that every capability writes into. It prevents CrawlKit from becoming a pile of scraping scripts by making every output inspectable, replayable, governable, and usable by other systems.
| Backbone object | Why it matters |
|---|---|
| Dataset / Asset | The durable product surface for rows, tables, views, artifacts, rates, mentions, pages, and generated data. |
| Schema / Contract | Structural and quality expectations that make data safe to query, serve, export, or train on. |
| Run / WorkflowRun / Job | The durable execution record for crawls, scrapes, ETL, pipelines, streams, synthetic runs, and domain pipelines. |
| Lineage | Source → job → transform → dataset → query/view/export/application provenance. |
| Quality result | Gate outputs before they reach production, reports, downstream apps, or AI training. |
| Policy / Consent / Audit | Governance for robots, source consent, sensitive fields, approvals, production changes, and accountability. |
| Query / Index / Flight endpoint | Consumption surfaces for SQL, semantic search, Arrow-native clients, exports, and application APIs. |
| Measurement / Telemetry | The learning loop: operational evidence, adoption analytics, attribution, calibration, and observability. |
The workbenches
Ingest
Crawls, Spider Studio, scrape jobs, frontier, stealth/proxies, WARC/Common Crawl, ETL, social sources, and domain sources.
Extract
Selectors, processors, extraction schemas, vertical templates, page handlers, content type detection, and deterministic records.
Refine
Enrichment, taxonomy, scoring, quality lab, dedup, decontamination, indexes, and measurement-ready signals.
Build
Workflow templates, pipeline DAGs, operation registry, schedules, AI plans, durable runs, worker progress, and lineage emission.
Query
SQL workbench, views, DataFusion, DuckDB hybrid, schema describe, stats, exports, indexes, and semantic search.
Serve
Exports, Arrow Flight, streams, training formats, reports, dataset cards, and downstream application APIs.
Quality
Contracts, gates, profiles, drift, deduplication, quarantine, taxonomy checks, and benchmark decontamination.
Governance
Policy, consent, robots/politeness, approvals, audit, sensitive fields, role scoping, and production safety.
Every object shows the Backbone Strip
Every capability detail page should answer the same evidence questions: What source produced this? Which run changed it? What dataset or asset did it create? What quality/policy gates passed? What lineage proves it?
Source / Definition → Output Dataset / Asset → Latest Run
→ Quality / Contract → Lineage / Impact → Policy / Audit
For example, a dataset detail view should show the spider or ETL run that created it, the latest run timestamp and row count, the quality gate pass/fail state, quarantined rows, upstream sources, downstream query views or applications, consent/policy status, and the audit trail for who changed what.
Applications and lenses
CrawlKit is not only a platform console. It also powers downstream applications and applied lenses.
| Layer | Examples | How it connects back |
|---|---|---|
| Downstream application | OpenRate healthcare pricing transparency | Every displayed rate can link to source MRF/TiC file, parser run, enrichment run, quality checks, query view, and audit. |
| SEO/GEO lens | Site audits, GSC, DNS, IndexNow, content briefs | Web crawl runs produce page datasets, issue datasets, content assets, and measurement feedback. |
| Social/revenue lens | Reddit/LinkedIn mentions, leads, campaigns, deals | Mentions become rows, enrichment/scoring creates leads, campaigns/deals trace back to source signals. |
| ML training-data lens | Synthetic data, quality filters, training formats, streams | Generated data keeps seed lineage, verification, consent, quality, decontamination, and export records. |
| Operations lens | Telemetry, alerts, healing, incidents, cost | Run/job failures and anomalies become evidence-backed incidents and safe remediation workflows. |
OpenRate is the proof point
OpenRate is the first downstream application built on the CrawlKit backbone. It demonstrates how difficult public data becomes a user-facing product.
CMS / hospital MRF / payer TiC sources
→ discovery + change detection + resilient fetch
→ parse + normalize + enrich + validate
→ hospital/provider/payer/rate datasets
→ query views for search, compare, discrepancies
→ reports, exports, APIs, and evidence drill-down
Where to go next
- Complete API Surface — all 383 operations across 54 domains.
- Arrow + DataFusion Guide — ETL, query, Parquet, exports, and serving.
- Workflows Guide — durable templates, operation registry, runs, schedules, and lineage.
- Spiders & Scraping Guide — AI once, deterministic forever.
- API Reference — typed request and response docs generated from the Rust API.