Outcrawl vs Scrapewise

Outcrawl 83

API reading. Scrape returns main content only by default with formats picking markdown, html, rawHtml, links, screenshots or schema JSON, search renders page bodies only on request, agent events take a level and an after cursor, the MCP server shows request bodies up to a character limit and says when it truncated, and the MCP surface is 29 tools with no toolsets (20 of 25). Cursor pagination on crawl results, agent results and downloads, limit, depth and path filters on crawl, status, outcome, task and age filters on sessions, and usage grouped by hour, day, month or capability (20). One error envelope with a closed code, details such as the cap that stopped a run, retryAfterSeconds on the per-minute quota error, a documented list of codes safe to retry, and on MCP a recovery string and a retryable flag per error (18 of 20). No idempotency key on any route, a repeated agent, crawl or monitor submit runs twice, and concurrency_limit carries no retry hint; cancel is idempotent, integration_changed (409) guards updates, and the npm MCP package sets readOnlyHint on GET tools, destructiveHint on DELETE tools and idempotentHint on either, which the hosted server was not read to confirm (10 of 20). Scrape needs a URL alone, agent runs need caps.budget on the API and SDKs with the MCP server defaulting it, and official SDKs exist in TypeScript and Python beside a CLI (15). 20+20+18+10+15 = 83.

Scrapewise 72

Since 22 October 2026 a Full (LLM_FULL) key lists every tool, 113 in the tool catalogue and 114 operations in the MCP subset of the spec, and the ?toolsets= address that kept the setup tools out of the default list is no longer documented, so more than 30 tools (5). A Read (LLM_READ) key lists only the read tools, a read-only subset of 48 (5 added back). Cursor pagination with limit up to 1,000, q, sort, from and to filters, and job results in pages of 100 (20). Error codes an agent can branch on, such as WALLET_INSUFFICIENT_BALANCE, scope_rejected, cost_exceeded, SCRAPER_RUN_ACTIVE and MCP_SHOP_LIMIT, with a troubleshooting table, though two envelope shapes are documented (16 of 20). Idempotency keys on paid tools and REST create and run, a same-name retry rule for list runs, and preview tokens before deletes. Annotations are documented but only some appear in the spec, and tools/list was not read (16 of 20). Few required fields and a free quote_only check, but only a Python SDK (10 of 15). 10 + 20 + 16 + 16 + 10 = 72.

Outcrawl 66

OAuth 2.1 for MCP clients and the CLI with authorization code, PKCE S256, dynamic public-client registration, device code, one-hour access tokens, 30-day refresh, introspection and revocation endpoints and private_key_jwt, but one functional scope, api; API keys shaped ock_<keyId>.<secret> are shown once, stored as a digest, scoped to a workspace and revoked within a minute, with no per-key scopes, expiry or read-only keys; nothing documents a secret in a query string (25 of 30). No read-only key or scope. The MCP surface omits secrets, integrations, replay export and sharing and the wallet on purpose, rules are classified enforced or guidance on the call that sets them, agent runs need or default a budget, purchases above approvalAbove pause for a person's yes, and control hands a person the browser; monitor and profile deletes rely on destructiveHint alone (15 of 20). Every tool returns untrusted page content. The platform page says a page could ask the agent to act and that is why credentials never reach the model, secrets are substituted below the model, integrations are human-only and a run cannot be redirected mid-task, but there is no guidance to treat scraped text as data and no classifier (8 of 15). Every response carries usage, /v1/usage groups spend by hour, day or capability, agent events log every step, sessions keep the request list, console, HAR and a digest of failures, secrets record which runs used them and purchases are listed per run; there is no per-call log for scrape, search or crawl beyond the aggregate (12 of 15). No security.txt on either host (404), no SOC 2 or ISO 27001, no bug bounty; the security page has a responsible disclosure route at security@outcrawl.ai with a 90-day window and a no-legal-action statement (6 of 20). 25+15+8+12+6 = 66.

Scrapewise 56

Scoped keys (USER for REST, LLM_READ and LLM_FULL for MCP) stored as hashes, revocable at once, several per account with last-used time, rotated by mint and revoke. No expiry and no finer scopes (26 of 30). LLM_READ cannot run paid tools or write, and deletes need a preview token committed within 5 minutes. Since 22 October the setup and destructive tools are no longer opt-in for a Full key, and LLM_FULL has no per-tool consent (16 of 20). A documented prompt-injection contract and guidance, but it rests on a type: "scraped" content block that the MCP specification of 25 November 2025 does not define, and the docs' statement that Anthropic's clients honour it by default was not confirmed (8 of 15). Every call is logged with a correlation id and key prefix and the wallet ledger itemises charges, but the operator sees logs only through support (6 of 15). No security.txt, disclosure policy, bug bounty or certification found, security.txt still 404 on 2 November (0 of 20).

Outcrawl 71

API reading, with the MCP server described. An OpenAPI 3.1 document at https://outcrawl.ai/docs/api-reference/openapi.json lists 62 operations over 51 paths with a summary each and the same 11 status codes, generated from the vendor's capability registry, but every request body is an untyped object marked additionalProperties: true with a pointer to the docs, every 200 response is an untyped object, and Error is the only schema; the typed shapes live in the SDK declarations and in the MCP server's zod schemas, which were read in the 0.4.3 npm package rather than from the hosted server (15 of 25). llms.txt and a 525 KB llms-full.txt on the docs, a Markdown twin of every page and an llms.txt on the site (10). The MCP descriptions each end with a Not for: clause naming the better tool and most with a Cost: clause, the OpenAPI summaries say when not to use a route, and the capability pages say which neighbour to use instead (18 of 20). Scrape formats are an enum of eight, event levels and record kinds are enums, caps.budget is required, and the SDK types bound every option, while the OpenAPI bodies carry none of it (10 of 15). curl, TypeScript and Python examples on every page, 20 worked programs, a 31-row error table with meanings and a retryable list, and MCP errors with a recovery string, though the closed union is sized 24, 31, 45 and 22 on different pages (13 of 15). /v1 in the path and 1.0.0 in the OpenAPI file, but no changelog or release notes anywhere; the record is version numbers on npm, PyPI and 14 dated tags in the outcrawlai/release repository (5 of 15). 15+10+18+10+13+5 = 71.

Scrapewise 79

Public OpenAPI 3.1 for the whole REST surface (247 operations on 2 November, 224 on 15 October) and an MCP subset (114 operations, 93 on 15 October), and MCP tools are generated from it with JSON Schema inputs (25). llms.txt, agents.md and pricing.md on scrapewise.ai, though docs.scrapewise.ai has no llms.txt (10). The data tools' descriptions say when to use each, what it returns, its price and that it is free or paid. Setup tools carry one-line descriptions (16 of 20). Required fields are marked and there are 55 enums in the MCP subset, but limits such as 1 to 20 results sit in prose, not as constraints, and paid tools wrap arguments in a body object (10 of 15). Request examples throughout, error codes listed with fixes, but the docs show two error envelope shapes and the MCP subset documents only 200, 204, 400 and 402 (10 of 15). A changelog with six entries, the newest items dated by day up to 23 October 2026, one evolving spec with no version prefix, and a one-release deprecation rule (8 of 15).