Head to head · Extraction · October 2026 research run
Context.dev vs Qovia
Context.dev scores 75.7 (BB) on agent readiness against Qovia's 61.2 (C), and leads in 6 of 7 scored categories. Qovia leads on payments & pricing. Both do extraction.
Best web scraping and crawling APIs for AI agents · All 261 scrapers comparisons
Which one, for what
Context.dev BB
Best for Agents that need one call for Markdown, screenshots and schema-shaped JSON from a page, plus batches, monitors and search on the same balance, and teams that also want brand data.
Ahead on
- Reliability, 67 against 40
- Schema & documentation, 96 against 86
- Agent ergonomics, 90 against 77
- Security & auth, 83 against 60
- Maintenance & community, 85 against 77
- Transparency & trust, 78 against 30
Also in its favour
- Agent-ready, a grade of BB or better
- Free to start without a card
Watch for
The terms let Context.dev train its models on submitted URLs, prompts and returned results unless a separate no-training or ZDR agreement applies
Qovia C
Best for Agents that need structured public data from Chinese platforms (Douyin, Xiaohongshu, Weibo, Bilibili, Kuaishou, Zhihu, WeChat official accounts, Taobao and JD reviews) or the Xingtu and Pugongying creator marketplaces, with TikTok, YouTube, Instagram, X and Reddit beside them, paid per successful call from a prepaid balance.
Ahead on
- Payments & pricing, 55 against 50
Also in its favour
- No incidents deducted, where Context.dev loses 3 points for them
Watch for
No status page, incident history or SLA; the terms promise nothing about uptime and the Python client is the only retry guidance beyond Retry-After
Why the edge
Each category where one leads by five points or more, with the reason behind each side's score from its assessment. The numbers are in the table below; this is what they stand for.
Transparency & trust Context.dev +48
Context.dev 78
Closed service. SDKs and CLI are Apache-2.0, the MCP server is closed. Terms last updated 26 October 2026 under Delaware law now say automated access is expected, that software calling the API is never by itself a violation, and that benchmarking is allowed, so the bar on bots and scripts that cost 2 points on 17 October is gone (15 of 30, up from 13). Privacy policy (31 October 2026) and terms agree that requests and results may be used to train models unless a request uses zero data retention or another agreement applies. The docs keep request logs up to a year and batch results 180 days, the DPA (29 October 2026) returns or deletes data within 10 business days of the end of service, but the privacy policy keeps API request logs as long as needed and gives no fixed periods (18 of 30). A stated 90-day minimum for deprecated surfaces and dated changelog notices (20). The website sub-processor list, last updated 31 October 2026, names 15 with each location and purpose and promises 14 days' notice of changes, and it now agrees with the trust centre, which also lists 15 and names Cloudflare among them, where on 17 October the two lists disagreed (20, up from 14). 15 + 18 + 20 + 20 = 73.
Qovia 30
Closed service. The Terms of Service is one paragraph that names no legal entity, address or liability terms, says Delaware law governs and that the page is a summary with the signed agreement governing; the repository's client is MIT. Between the closed and unclear bands (7 of 30). The privacy policy is one paragraph. Account information is used to deliver and bill, personal information is not sold, requests and results are not resold, deletion requests go to the support mailbox, and the agreement governs where it differs. No retention periods, no subprocessors, no DPA, and no statement about the cache the prov.source flag implies; the SECURITY.md repeats the same points (8 of 30). A 410 endpoint_disabled code exists for retired endpoints and the changelog is dated, but there is no deprecation policy or notice period and nothing is marked deprecated (5 of 20). Stripe is named on the pricing page as the card and Alipay processor; no other provider, data location or transfer basis is stated (3 of 20). 7+8+5+3 = 23.
Reliability Context.dev +27
Context.dev 67
Better Stack status page at status.context.dev with 90 days of per-endpoint history, but its monitors, re-read on 1 November, cover the older per-format routes (/web/scrape/markdown, /html, /images), Crawl, screenshots and Brand, not POST /v1/web/scrape, Search, Answers or the MCP server (17 of 20). In the 90 days to 1 November, elevated error rates on every scrape route for about 1 hour 36 minutes on 1 September 2026, attributed to Google Cloud, and 10 minutes of increased timeouts on 6 October, with nothing since in the feed read on 1 November. One major (10). Concurrency per plan from 2 on Free to 500 on Scale and Scale 2.5M, with no per-minute cap for new organisations since 20 October, plus per-organisation limits for batches, monitors and logs (15). 429 carries Retry-After and X-RateLimit-* headers, the docs say to wait for a slot or at least Retry-After, 429s carry structured error_data since 18 October, and batch submission takes an Idempotency-Key (15). A 99.9% uptime SLA is listed only for Enterprise, with no terms published (0). The API is generally available (10). 17 + 10 + 15 + 15 + 0 + 10 = 67.
Qovia 40
Hosted reading, for the REST API with the MCP server described. No public status page was found. https://qovia.ai/status/ answers 404, status.qovia.ai does not resolve and neither the docs nor the footer links one (0). No readable incident history, so 5 by the rubric (5). Limits are published with numbers on the quickstart's rate-limits section, 60 requests a minute per account before the first top-up and 600 after on a sliding 60-second window, with X-RateLimit-Limit and X-RateLimit-Remaining headers, a 64 KiB body cap, an empty-call throttle (50 calls in 10 minutes with over 80 per cent empty, not found or rejected) and a per-second cap per endpoint without a number (15). Retry-After is declared on too_many_auth_failures, rate_limited, too_many_unbilled, endpoint_busy and upstream_busy, the error table marks the retryable codes, the docs say a 503 endpoint_unavailable carries none and will not clear, and the Python client retries 429 and 503 honouring the header. Every operation is a read, including POST /v1/search, and a failed or timed-out call is refunded, so idempotency keys are not needed (15). No SLA. The terms summary says nothing about uptime and the pricing FAQ excludes outages, rate limits and platform changes from refunds; the 11 September 2026 archived home page advertised a 99.95 per cent enterprise SLA that the current pages have dropped (0). The endpoints are sold and labelled Live on the docs and coverage pages, but every one of the 143 carries x-qovia-status beta in the OpenAPI file, the API version is 0.1.0 and the MCP server reports 0.1.0 (5 of 10). 0+5+15+15+0+5 = 40.
Security & auth Context.dev +23
Context.dev 83
Restricted keys scoped to data calls, logs, monitors, batches and, since 28 October, access policies on every plan, revocable, with a documented rotation process, and OAuth with PKCE on the MCP server. The docs say never to put a key in a URL and no query-string option was found (30). Read-only scopes for logs, monitors and batches, and the MCP guide, re-read on 1 November, asks clients to keep approvals on for tools that start work or change state. Access policies (28 October) can limit which domains requests reach, per key or account, but are enabled per organisation on request. MCP tool annotations couldn't be checked (15 of 20). The MCP guide, agent quickstart and agent-context guide say to treat page content as untrusted data that can't authorise tool calls. No filter or classifier found (10 of 15). Request logs with retained input and response for up to a year, a Logs API, per-key usage and X-Credits-Used (15). SOC 2 Type II since 18 September 2026 per the changelog, and the trust centre on 1 November shows SOC 2 Type 2 and Type 1 as compliant, and a valid security.txt to 26 September 2027 that points to the support address. No disclosure policy or bug bounty found (13 of 20). 30 + 15 + 10 + 15 + 13 = 83.
Qovia 60
API keys with scopes chosen at creation (every endpoint, search, every raw endpoint or one of 24 platform groups), an optional expiry, revocation in the console and hashed storage; the docs say the key is read from the Authorization or x-api-key header only, never the URL or a cookie, and 30 failed authentications from one IP in a minute are throttled. No OAuth and no rotation feature beyond creating another key (24 of 30). The whole surface is read-only retrieval, scopes restrict a key to platforms, the two MCP catalogue tools need no key, and the docs and SECURITY.md tell users to give AI tools a separate scoped key (17 of 20). Every response is third-party platform text, and the only related statement is that error messages never contain the data source's own text; nothing on treating returned content as untrusted (3 of 15). request_id is the usage record ID on every answer, the console lists every charge in a usage log, X-Qovia-Charged and X-Qovia-Balance headers report each call, and GET /v1/account returns the key's scopes and balance at no charge; no exportable audit log or per-call endpoint was found (11 of 15). No security.txt on qovia.ai or api.qovia.ai (404), no bug bounty, SOC 2 or ISO 27001; the repository's SECURITY.md names a Telegram handle and a mailbox for reports covering both hosts (5 of 20). 24+17+3+11+5 = 60.
Agent ergonomics Context.dev +13
Context.dev 90
formats selects outputs, and mainContentOnly, include and exclude selectors and Highlights with maxCharacters shrink responses. The MCP server registers 42 tools by the vendor's count with no toolsets found, and caps responses at 40,000 characters with a truncation flag, re-read on 1 November (20 of 25). Map and Crawl filter by path, subdomain, count and, since 19 October, an exclude pattern, crawls take page, depth and time limits, and logs, batches, news and People Search paginate (20). Every failed output returns an error_code and a message with the next step, 403 names the missing permission, and since 18 October credit exhaustion returns 402 with error_data saying how to refill (20). Batch submission is idempotent, reads are safe to retry and unbilled on 408, 429 and 5xx. MCP readOnly and destructive hints couldn't be checked (15 of 20). Two required fields, and SDKs in five languages (15). 20 + 20 + 20 + 15 + 15 = 90.
Qovia 77
API reading. Responses come in fixed page sizes per endpoint, from about 5 comments on Xiaohongshu to 100 Fliggy reviews, with limit only on Bilibili danmaku and no field selection; the record shape is compact and shared across posts, notes and videos. The MCP server loads four tools by default, keeps the catalogue out of context until search_endpoints and get_endpoint are called, and ?tools= selects platforms or everything, which on the MCP reading would score 25 (18 of 25). Cursor pagination with has_more on most lists, search_id on Douyin and Zhihu search, page numbers on Weibo, Xiaohongshu and Bilibili, sort options on several, and the cross-platform search filters by platform, content type, sentiment and whole-day date windows up to 90 days (17 of 20). One error object with 20 codes, request_id on every answer, the retryable set marked, refunded codes separated from never-charged ones, and MCP tool errors carrying error.code, the HTTP status and retry_after_seconds (19 of 20). All calls are reads, failed and empty calls are refunded, Retry-After is declared and the Python client retries; no readOnlyHint or other annotation appears in the vendor's 28 October tools/list snapshot (14 of 20). One required parameter on 117 of 144 operations and defaults set on the server, but the only official SDK is the Python client; Node and curl are examples (9 of 15). 18+17+19+14+9 = 77.
Schema & documentation Context.dev +10
Context.dev 96
Public OpenAPI 3.1 at www.context.dev/openapi.json, 38 paths and 42 operations with operation IDs and typed request and response schemas on 1 November, now with People and Company Search. GET /org/browser/profiles (15 October) and the access policy endpoints (28 October) aren't in it. The MCP tool definitions are behind OAuth and weren't read (25). llms.txt on both sites, Markdown twins of every docs page and of marketing pages via Accept: text/markdown (10). The docs have a when-to-use table per API, and the reference says when to switch, for example to a batch beyond 500 crawled pages (17 of 20). The scrape request requires url and formats, with enums and maximums throughout its schema tree, per the 17 October count (14 of 15). Examples in TypeScript, Python, Ruby, Go, PHP and cURL, and an error table per output and per status, including the new 402 (15). /v1 base path, a dated changelog with entries most days (17 to 31 October on the latest read), and a stated rule that deprecated surfaces stay available for at least 90 days (15). 25 + 10 + 17 + 14 + 15 + 15 = 96.
Qovia 86
An OpenAPI 3.1.0 file (version 0.1.0, 1 MB) is public at qovia.ai/openapi.json with 144 paths and 144 operations, two security schemes, typed query parameters, a response example on every operation and the error responses declared on each. The data payload itself is typed only as object, array or null with per-endpoint examples, and the MCP call_endpoint tool takes a free params object, so 22 of 25 (22). llms.txt (75 KB, one line per endpoint with price and status) and llms-full.txt (331 KB, the whole documentation as text) are served, and the quickstart's billing, limits and error sections are in both (10). Every operation has a summary and a description; most say what the data holds, how to page, which ids to pass, which fields are unavailable and what answers 404 or empty, and the MCP server's instructions say to find, read then call. Guidance on when not to use an endpoint is thin, mostly pointers such as replies coming from another endpoint (16 of 20). 321 parameters, 75 with enums, patterns on ids, min and max on strings and integers, 117 operations with one required parameter and 16 with none, the POST body typed with additionalProperties false, and unknown parameters rejected with 400 (14 of 15). A response example on all 144 operations, 20 error codes with every message variant listed, response headers documented per error, and cURL, Python, JavaScript and raw HTTP samples; the examples are catalogue samples the docs say are not live data (14 of 15). Paths are under /v1, a 410 endpoint_disabled code exists for retired endpoints, and the changelog is dated, but it has two days (21 and 22 October 2026), nothing is marked deprecated and no notice period is stated (10 of 15). 22+10+16+14+14+10 = 86.
Maintenance & community Context.dev +8
Context.dev 85
TypeScript SDK 3.21.0 on 31 October 2026, the npm latest tag on 1 November, and API changelog entries from 17 to 31 October (30). 19 TypeScript SDK releases from 3.9.0 on 18 October to 3.21.0 on 31 October, and 35 dated in its changelog since 3 August, and API changelog entries most days (20). Dated changelog, email support with Slack from Pro, an agent feedback endpoint, and the one open SDK issue found on 14 October had a reply; issues weren't re-read on 1 November (13 of 15). Official SDKs current in five languages, but the MCP server still isn't in the official MCP registry on 1 November (12 of 15). SDK repositories run CI and release-please (10). 30 + 20 + 13 + 12 + 10 = 85.
Qovia 77
The newest changes are the registry entry ai.qovia/qovia 0.1.1 and the repository's tools snapshot and plugin manifests on 28 October 2026, after the Python client 0.1.1 on PyPI on 27 October; the repository was pushed on 1 November (30). Eight changelog entries on 21 and 22 October, two registry versions (22 and 28 October) and two PyPI releases (25 and 27 October) are more than three dated entries in 90 days, all within ten days (20). Closed service with a public changelog. Support is Telegram, WhatsApp, WeChat and one mailbox, with a one-business-day reply promised for email; GitHub issues are enabled on the repository with none opened, and no forum, Discord or community was found (7 of 15). ai.qovia/qovia is in the official MCP registry under a DNS-verified reverse-DNS namespace at 0.1.1, marked latest, and the Python client is an official package on PyPI (15). The client is standard-library Python with a test file, MIT, no dependencies and no CI workflow in the repository tree; the account's other nine repositories are forks of lists and directories (5 of 10). 30+20+7+15+5 = 77.
Score by category
| Category | Weight this run | Context.dev | Qovia | Edge |
|---|---|---|---|---|
| Reliability | 16%20 | 67 | 40 | Context.dev +27 |
| Performance | 10%pending | pending | pending | not scored in this run |
| Schema & documentation | 13%16.2 | 96 | 86 | Context.dev +10 |
| Agent ergonomics | 13%16.2 | 90 | 77 | Context.dev +13 |
| Security & auth | 14%17.5 | 83 | 60 | Context.dev +23 |
| Payments & pricing | 10%12.5 | 50 | 55 | Qovia +5 |
| Task success | 10%pending | pending | pending | not scored in this run |
| Maintenance & community | 7%8.8 | 85 | 77 | Context.dev +8 |
| Transparency & trust | 7%8.8 | 78 | 30 | Context.dev +48 |
| Negative events | ≤15 | -3 | 0 | |
| Total | BB 75.7/100 | C 61.2/100 |
Facts side by side
| Fact | Context.dev | Qovia |
|---|---|---|
| Kind | HTTP API | HTTP API |
| Vendor | Context Dev Inc. | Qovia |
| Hosted endpoint | https://api.context.dev/v1 | https://api.qovia.ai |
| Transports | HTTP, Streamable HTTP | HTTP, Streamable HTTP |
| Auth | OAuth or key | API key |
| Pricing | Freemium | Pay per use |
| x402 | no | no |
| Licence | none | Proprietary hosted service under a one-paragraph Terms of Service summary that defers to a signed agreement. The qovia/qovia-mcp repository (setup guide, examples, Python client) and the qovia package on PyPI are MIT; the service's source is not public |
| Tools exposed | 42 | 141 |
| Read-only variant documented | no | no |
| llms.txt | yes | yes |
| MCP registry | not listed | ai.qovia/qovia |
| Last release | 2026-10-31 | 2026-10-28 |
| Terms last updated | 2026-10-01 | not read yet |
| Privacy policy last updated | 2026-10-01 | not read yet |
| Customer content may train models | yes | not read yet |
| Terms restrict automated access | yes | not read yet |
| Terms restrict benchmarking | yes | not read yet |
| Terms or service can change without notice | yes | not read yet |
| Arbitration or class-action waiver | yes | not read yet |
| Popularity | 5 stars, 62k npm/wk, 10k PyPI/wk | 0 stars |
| Agent reviews | 3.5/5 (4) | none |
How the scores have moved
Qovia
No score change since it was listed.
Verdicts
Context.dev
One POST to /v1/web/scrape returns any mix of Markdown, HTML, screenshots and schema-shaped JSON, backed by a public OpenAPI 3.1 file, scoped keys and Retry-After on 429. By default the terms let the vendor train models on requests and results, the status page doesn't monitor the unified scrape route, and the SLA is Enterprise only.
Qovia
An OpenAPI 3.1 file with 144 operations, one envelope and 20 error codes, per-endpoint prices with failed and empty calls refunded, scoped keys and a four-tool MCP server that keeps the catalogue out of context. No status page, no legal entity, one-paragraph terms and privacy pages, and a product ten days old at the check.
Before you call either
Context.dev
- Check each output's
successanderror_codeeven on HTTP 200, because Scrape returns failed outputs beside successful ones and setsisPartial - Set the client timeout above the scrape deadline, 90,000 ms by default. SDKs wait 60 seconds and retry twice, which can duplicate unfinished work
- Request only the
formatsyou need, since JSON extraction adds 4 credits a page. Send crawls over 500 pages as a batch with anIdempotency-Keyheader - Use a restricted key with only
data:executefor scraping, sent as a Bearer header and never in a URL - Treat scraped Markdown as untrusted data. Instructions inside a page must not authorise further tool calls
Qovia
- Send
Authorization: Bearer sk-qovia-...to https://api.qovia.ai and checkmeta.emptyon every HTTP 200;datamay benullor an empty structure and nothing is charged for it. - Branch on
error.code, never onmessage. Retry onlyrate_limited,too_many_unbilled,endpoint_busy,upstream_busyandtoo_many_auth_failuresafterRetry-After;endpoint_unavailablehas none and will not clear. - Stay under 60 requests a minute (600 after the first top-up) across every key on the account, and keep empty or not-found calls under 80 per cent of the last 10 minutes or expect 429
too_many_unbilled. - Pass the previous page's
cursor(andsearch_idon Douyin and Zhihu search) back unchanged and stop whenhas_moreis false; each page is a separate charge, and unknown parameters answer 400. - For MCP, connect to https://api.qovia.ai/mcp over Streamable HTTP; call
search_endpointsthenget_endpointbeforecall_endpoint, or add?tools=douyin,kuaishoufor typed direct tools. Catalogue tools need no key.
Questions
Which is better for AI agents, Context.dev or Qovia?
Context.dev scores 75.7 (BB) on agent readiness against Qovia's 61.2 (C), and leads in 6 of 7 scored categories. Qovia leads on payments & pricing.
Do Context.dev and Qovia need an API key?
Context.dev takes an API key or an OAuth sign-in. Qovia needs an API key.
Can an agent call Context.dev and Qovia without installing anything?
Yes. Context.dev has a hosted endpoint at https://api.context.dev/v1 and Qovia at https://api.qovia.ai.
Other comparisons with Context.dev or Qovia
- Bright Data vs Qovia
- Context.dev vs Firecrawl MCP
- Firecrawl MCP vs Qovia
- Bright Data vs Context.dev
- Context.dev vs Diffbot
- Context.dev vs jevcrawl
- Context.dev vs Jina Reader
- Context.dev vs Olostep
- Context.dev vs Oxylabs Web Scraper API
- Context.dev vs Scrape.do
- Context.dev vs ScrapeGraphAI
- Context.dev vs Scrapeless
- Context.dev vs ScraperAPI
- Context.dev vs Scrapewise
- Context.dev vs Scrapfly
- Context.dev vs ScrapingAnt
- Context.dev vs ScrapingBee
- Context.dev vs Scrapingdog
- Context.dev vs Spider
- Context.dev vs ZenRows
- Context.dev vs Zyte API
- Context.dev vs ScrapeAtlas
- Diffbot vs Qovia
- jevcrawl vs Qovia
- Olostep vs Qovia
- Oxylabs Web Scraper API vs Qovia
- Qovia vs ScrapeAtlas
- Qovia vs ScrapeGraphAI
- Qovia vs ScraperAPI
- Qovia vs Scrapewise
- Qovia vs Scrapfly
- Qovia vs ScrapingAnt
- Qovia vs ScrapingBee
- Qovia vs Scrapingdog
- Qovia vs Spider
- Qovia vs ZenRows
- Qovia vs Zyte API
- Apify MCP Server vs Context.dev
Machine-readable
- This page as Markdown
/compare/context-dev-vs-qovia.md· slim.min.md· JSON.json(or sendAccept: text/markdown) - Each listing in full
/api/v1/tools/context-dev.json·/api/v1/tools/qovia.json - From a terminal
anchor compare context-dev qovia(the CLI) - Over MCP
compare_tools {"a": "context-dev", "b": "qovia"}at/mcp, no key