Crawl4AI vs Firecrawl MCP

Crawl4AI 35

Graded on the hosted lines, for Crawl4AI Cloud, which the README and llms.txt now lead with and which an agent calls through its MCP server; the library would read on the local lines. No status page is linked, and status.crawl4ai.com fails TLS (0). No incident history readable (5). Per-plan limits are published with numbers in /v1/prices, from 10 requests a minute and 2 at a time on Free to 1,500 and 50 on Scale (15). 429 carries X-RateLimit headers and a message, 503 busy says to resend after a few seconds, and MCP job submissions take an idempotency key, but no Retry-After or backoff guidance (10 of 15). No SLA found in the terms or pricing (0). Prices and the free credit are a soft launch, /serp is beta and /answer experimental, while scrape, extract and search carry no beta label (5 of 10). 0 + 5 + 15 + 10 + 0 + 5 = 35.

Firecrawl MCP 78

Better Stack status page at status.firecrawl.dev with history (20). Four incidents in the 90 days to 1 November 2026, all partial degradations of api.firecrawl.dev, the longest 56 minutes on the /interact prompt field on 27 August (06:00 to 06:56 UTC in the feed) and the others 7, 16 and 46 minutes, with none after 15 September. The 20 July /interact incident has left the window (20). Rate limits published per plan and endpoint, from 10 scrapes a minute on Free to 10,000 on Scale, plus concurrent browsers and a per-IP daily cap for keyless use, re-read on 1 November (15). The Errors reference, linked from the rate-limit page, says 429s carry Retry-After when available, marks every error retryable or not, and gives backoff advice per code. No idempotency keys for calls that start crawl, agent or monitor jobs (13 of 15). SLA on Enterprise only, per the pricing page, with no terms published (0). GA (10). 20 + 20 + 15 + 13 + 0 + 10 = 78.

Crawl4AI 55

No OpenAPI or similar contract for the cloud API, and the MCP tool schemas are only in the live tools/list, which needs a key. /v1/prices is machine-readable but covers prices, not calls (0). llms.txt of 304 lines written for agents (10). Each endpoint says what it is for and when to choose another, such as trying without a proxy first, using browser only when auto misses content and /scrape/jobs for slow lists (16 of 20). Parameters are listed with types, defaults, allowed values (format both, md or html; proxy none, isp or residential) and limits, in prose (10 of 15). Curl examples per endpoint and documented error strings per status with whether each is charged (14 of 15). No changelog for the cloud API, only dated notes such as /serp 'added 8 Oct 2026'. The library has a dated CHANGELOG (5 of 15). 0 + 10 + 16 + 10 + 14 + 5 = 55.

Firecrawl MCP 91

Public OpenAPI at docs.firecrawl.dev/api-reference/v2-openapi.json (and v1), and zod schemas on every MCP tool (25). llms.txt and Markdown docs (10). The README and server instructions say when not to use a tool, for example to look elsewhere when a browser session must be driven step by step across many calls, and 3.33.4 added when-to-use sentences to the search and scrape descriptions (18 of 20). Typed inputs with enums, but open issue #325 counts 132 parameters with no description and #373 reports a published schema that disagrees with the API, per the 1 October check, not re-read on 1 November. 3.35.0 briefly advertised record-typed inputs as closed empty objects, fixed in 3.36.0 the next day (11 of 15). Examples throughout, keyless failures return structured recovery payloads with next_actions and signup_url, and an Errors reference lists each HTTP status and error string with cause, remedy and whether to retry (15 of 15). A CHANGELOG and firecrawl.dev/changelog. The CHANGELOG now has dated entries for 3.29.0 to 3.36.0, but still skips 3.22 to 3.24 and 3.26 to 3.28 and lists 3.25.0 as unreleased (12 of 15).

Crawl4AI 67

16 MCP tools, six of them for billing and support (15 of 25). format md or html, parse opt-in blocks, batch limits and job results paged 500 at a time over HTTP and 50 over MCP (16 of 20). String error codes an agent can branch on (blocked, login-wall, no-answer, busy, no_credit, plan_cap, spend_cap), each with whether it is charged (18 of 20). Failed and cancelled calls aren't charged and jobs_submit takes an idempotency key, but no tool annotations are documented and two tools can charge the saved card (10 of 20). Only url or q is required. There is no official SDK for the cloud API; the docs give curl and standard-library Python and Node examples (8 of 15). 15 + 16 + 18 + 10 + 8 = 67.

Firecrawl MCP 92

27 tools in the full profile by the README's count (28 on the hosted endpoint with a key, per the contract snapshot), 3 on the keyless endpoint and 9 on the search-only endpoint, and env flags drop the feedback tools (23 of 25, the 11 to 30 band plus 8 for fixed subsets). Format selection, main-content filtering, page limits on map and crawl, and results over 20,000 estimated tokens handed off to retained storage instead of inline (20). Recovery payloads with next actions, and since 3.35.0 API refusals carry a machine-readable code, status and retryAfterSeconds in structuredContent, but schema mismatches reported in open issues (17 of 20). readOnlyHint or destructiveHint on all 28 tools of the hosted full profile, with monitor delete and interact stop marked destructive, and since 3.35.0 idempotentHint on the crawl, agent, interact and monitor tools. No idempotency keys for crawl or agent jobs (17 of 20). One required argument on most tools, and official SDKs in Python and Node (15).

Crawl4AI 46

Bearer keys, shown once and regenerable, with 3 to 30 keys per plan. claude.ai and ChatGPT sign in through a flow that advertises one scope, mcp, and each app shows as a revocable key. No per-key scopes, and the docs don't offer the key in a URL (22 of 30). A key can call billing routes, and the MCP topup and recharge tools charge the saved card, though llms.txt tells agents to ask first and accounts can set a spend cap. No read-only key (6 of 20). Scraped pages are untrusted text and no prompt-injection guidance was found in llms.txt or the API docs (0). Per-call cost, balance and fetch path in response headers and fields, no access log documented (6 of 15). The security page (4 October 2026) gives a disclosure process with fix targets and says SOC 2 is in progress, and the library publishes GitHub advisories with reporter credits. security.txt redirects to a 404, and no bounty found (12 of 20). 22 + 6 + 0 + 6 + 12 = 46.

Firecrawl MCP 70

OAuth on the hosted endpoint /v2/mcp-oauth, a separate OAuth identity for the search-only endpoint, and plain revocable keys as Authorization: Bearer. Keys locked to endpoints and formats, and IP allowlists, are Enterprise only (25 of 30). The README says never to put a key in the server URL. 3.29.1 removed the credential-in-path route from the repository's service image, and it was never documented as an option, so no deduction. 3.36.0 checks Host and Origin in local HTTP mode against DNS rebinding. Keyless and search-only surfaces limit what a session can do, write tools carry destructiveHint, and Alexandria provider terms need explicit user consent and confirmed: true (18 of 20). Retained Alexandria results tell the model that source content is data, not instructions, and Enterprise Threat Protection blocks risky URLs, off by default. The API's opt-in checkPromptInjection classifier is now a top-level scrape option for any format, markdown included, and for crawl, search and monitors (+4 credits, 403 SCRAPE_PROMPT_INJECTION_DETECTED). The MCP tool schemas still don't expose it, so an agent using this server can't turn it on (8 of 15). firecrawl_credit_usage shows spend, by endpoint and by day since 3.34.0, but its description says it can't attribute single jobs, URLs or MCP clients, and we found no per-call log for operators (5 of 15). SOC 2 Type II on the Enterprise page, and security.txt valid to 10 June 2027 with a Bugcrowd vulnerability disclosure programme, re-read on 1 November. CVE-2026-116066 (High in GitHub's unreviewed database, published 25 October), a DNS-rebinding SSRF in the Firecrawl API's Playwright service before 2.11.586, was fixed in a public commit on 21 October that cites issue #5256. No bug bounty found (14 of 20).

Crawl4AI 45

No x402, MPP or L402 (0). Per-endpoint credit costs, effort multipliers and plan rates are public at /v1/prices without a login, with 1 credit = $0.001 (20). 1,000 free credits a month, no card (20). A 7-day pass key comes from a click on api.crawl4ai.com without an account, which is still a person in a browser, and a permanent key needs email verification (5 of 20). 0 + 20 + 20 + 5 = 45.

Firecrawl MCP 50

No x402, MPP or L402 in the README, pricing page or agent onboarding guide, re-read on 1 November (0). Plans are public with per-endpoint credit costs (1 a page, search 2 per 10 results, +4 for JSON formats), and $5 top-ups only on paid plans, so between plan-only and per-unit (15 of 20). Free plan with 1,000 credits a month, no card (20). The hosted keyless endpoint gives scrape, search and parse with no signup, but crawl, map and agent need a key. The agent onboarding guide adds key routes through Stripe Projects and, on supported platforms, a WorkOS ID-JAG assertion, both of which start from a person's existing login (15 of 20).

Crawl4AI 70

The library is Apache-2.0 and the cloud is closed under terms that name CONTEXT4AI PTE LTD (22 of 30). The privacy policy gives retention periods (account plus 90 days, usage records 13 months, logs 90 days), but none for crawled content, and llms.txt says repeated URLs are served from a shared archive while the security page says fetched pages are used only to answer the request (18 of 30). The legacy cloud API closed on 6 November 2026 with a migration guide but no dated notice found, and the terms promise notice of material changes only 'where reasonable' (8 of 20). Subprocessors named with their locations on the security page, with EU storage of account records (18 of 20). 22 + 18 + 8 + 18 = 66.

Firecrawl MCP 75

MIT (30). Zero data retention on Enterprise and a DPA from Standard up, but the privacy policy, re-read on 1 November, still dates from 26 December 2024, keeps personal data until you ask for deletion and gives no retention period for scraped content (12 of 30). The v1 endpoints are kept as legacy alongside v2 and the pricing change of 4 September 2026 was dated, but we found no deprecation policy (8 of 20). Data stored in the United States, and Stripe, PostHog, Crisp and Vercel Analytics are named, with no full subprocessor list (10 of 20).