fastCRW, in verified numbers
fastCRW is a Firecrawl-compatible web scraping, crawling, and search API for AI agents. The engine is a single static Rust binary — no Redis, no Node.js, no containers required — released open source under AGPL-3.0. The same API runs self-hosted for free or on the managed cloud, so there is no exit cost.
On Firecrawl's own public 1,000-URL dataset, fastCRW has the highest truth-recall of three tools tested — 63.74% vs Crawl4AI 59.95% and Firecrawl 56.04% — at a median scrape latency of 1914 ms. The numbers below trace to the public OSS README and the reproducible result of record.
Product identity
| Product | fastCRW — web scraping, crawling & search API for AI agents |
| Engine | Single static Rust binary (no Redis, Node.js, or containers required) |
| License | AGPL-3.0 (open source) |
| Public repository | github.com/us/crw |
| MCP package | crw-mcp on npm (dist-tag latest: 0.6.0) |
| API style | Firecrawl-compatible REST — drop-in after a base-URL swap |
| Managed cloud | fastcrw.com |
Accuracy and latency, 3-way
fastCRW returned the most accurate content of the three tools tested. Source: Firecrawl's public scrape-content-dataset-v1 (1,000 URLs, 819 labeled), scored by the open diagnose_3way.py harness — a single run of 3,000 requests dated 2026-05-08.
| Metric | fastCRW | Crawl4AI | Firecrawl |
|---|---|---|---|
| Truth-recall (of 819 labeled URLs) | 63.74% | 59.95% | 56.04% |
| Scrape-success (of 1,000 URLs) | 87.7% | 83.5% | 89.7% |
| Thrown errors (of 3,000 requests) | 0 | 0 | 0 |
| p50 latency | 1914 ms | 1916 ms | 2305 ms |
| p90 latency | 14157 ms | 4754 ms | 6937 ms |
| p99 latency | 15012 ms | 13749 ms | 21107 ms |
The p90 tail is disclosed, not hidden. fastCRW's p90 of 14157 ms is the slowest of the three — the cost of a chrome-stealth fallback that retries hard pages instead of giving up. That same retry is what lifts truth-recall to first place: the slow tail and the accuracy lead are the same mechanism. See the full methodology on the benchmark page.
Search latency, 3-way
A separate 100-query benchmark run concurrently against all three providers. fastCRW won 73 of 100 latency races and averaged 880 ms across the run.
| Metric | fastCRW | Firecrawl | Tavily |
|---|---|---|---|
| Average latency | 880 ms | 954 ms | 2,000 ms |
| Median latency | 785 ms | 932 ms | 1,724 ms |
| P95 latency | 1,433 ms | 1,343 ms | 3,534 ms |
| Latency wins (of 100 queries) | 73 | 25 | 2 |
Plans and credits
Self-hosting the AGPL-3.0 engine is free — $0 per 1,000 scrapes versus $0.83–5.33 on Firecrawl's hosted tiers. Managed cloud launch pricing reverts to regular price on 2026-06-01.
| Plan | Launch price | Regular price | Credits |
|---|---|---|---|
| Free | $0 | $0 | 500 one-time lifetime credits |
| Hobby | $13/mo | $19/mo | 3,000 / mo |
| Standard | $69/mo | $99/mo | 100,000 / mo |
| Growth | $279/mo | $399/mo | 500,000 / mo |
| Scale | $549/mo | $749/mo | 1,000,000 / mo |
Credit costs: scrape 1 (chrome render 2), crawl 1/page, search 1/query, map 1, extract 5, browse 1/session. See live pricing.
What it costs to run
These are structural facts, not benchmark claims. fastCRW ships as a single ~8 MB binary and needs 1 container (plus an optional sidecar); Firecrawl's stack is ~2–3 GB across 5 containers. The opt-in Chrome Docker Compose variant is ~500 MB image and ~1 GB resident.
| Metric | fastCRW | Firecrawl |
|---|---|---|
| Docker image | single ~8 MB binary | ~2–3 GB total |
| Containers needed | 1 (+ optional sidecar) | 5 |
What fastCRW does not do
An accurate fact sheet states its gaps. fastCRW is a focused scraping API, not a managed-browser platform:
- Screenshot output — a request for formats: ["screenshot"] returns HTTP 422.
- Multi-URL batched /v1/extract — iterate /v1/scrape or use /v1/crawl instead.
- Agent (Spark) models, deep-research endpoint, and Fire-engine anti-bot.
- Persistent session or state — fastCRW is stateless per request.
- LLM extraction beyond OpenAI and Anthropic providers.
robots.txt is respected by default and may be overridden only when the caller has the legal right to do so.
Common questions
What is fastCRW?
fastCRW is a Firecrawl-compatible web scraping, crawling, and search API for AI agents. The engine is a single static Rust binary — no Redis, no Node.js, no containers required — released open source under AGPL-3.0. The same API runs self-hosted or on the managed cloud, so there is no exit cost.
How accurate is fastCRW compared to Firecrawl and Crawl4AI?
On Firecrawl's own public 1,000-URL scrape-content dataset, fastCRW recovered the labeled content on 63.74% of 819 matchable URLs — the highest of three tools tested, ahead of Crawl4AI (59.95%) and Firecrawl (56.04%). The run is reproducible with the open diagnose_3way.py harness.
How fast is fastCRW?
On the same 1,000-URL benchmark, fastCRW's median (p50) scrape latency is 1914 ms, against 1916 ms for Crawl4AI and 2305 ms for Firecrawl. Its p90 of 14157 ms is the highest of the three: the disclosed cost of a chrome-stealth fallback that recovers hard URLs the others drop.
How much does fastCRW cost?
Self-hosting the AGPL-3.0 engine is free — $0 per 1,000 scrapes, versus $0.83–5.33 on Firecrawl's hosted tiers. Managed cloud plans start at a $0 Free tier (500 lifetime credits) and a $13/mo Hobby plan; the Standard plan is $69/mo for 100,000 credits during launch pricing.
Is fastCRW a drop-in Firecrawl replacement?
On the overlap surface — /v1/scrape, /v1/crawl, /v1/map, /v1/search — request shapes match closely, so most migrations are a base-URL swap plus minor adjustments. Response field names and error envelopes have small divergences, so it is API-compatible rather than byte-identical.
What does fastCRW not support?
fastCRW does not return screenshots (a screenshot request returns HTTP 422), and it has no multi-URL batched extract, no agent or deep-research endpoints, and no persistent sessions. LLM extraction supports OpenAI and Anthropic providers. These gaps are stated plainly rather than hidden.