Apify MCP Server

The server can narrow exposed capabilities through an explicit tools parameter, tools carry readOnlyHint and openWorldHint annotations, and defaults plus auto-injection are documented, so least privilege and external effects are handled reasonably well. The default nevertheless includes call-actor and a web-browsing Actor, while dynamic discovery can broaden effective capability. Payment flow, prepaid amounts, refund conditions, external Actor execution, and result-storage steps are disclosed, but there is no evidence that every costly or side-effecting call requires user confirmation; direct x402 may automatically sign a $1 transaction through the client. Tokens are supplied through environment variables or Authorization headers, CI masks raw and URL-encoded tokens, and tests remove an mcpc envelope containing plaintext credentials before assertions or failures. The supplied material does not show comprehensive retention, log-redaction, or telemetry-field policies. Dependencies and CI checks are visible, but no lockfile, vulnerability scan, SBOM, or dependency-update policy is supplied. Actor runs can be aborted and some payment paths refund unused balances, but there is no general transactional rollback. Repository, package author, issue tracker, and copyright attribution are present, although publisher identity remains unverified as stipulated.

The README, package metadata, CI, and test helpers are highly consistent about the Node 22 minimum, package identity, transports, tool configuration, token requirements, and telemetry being disabled by default in tests. CI runs unit tests on Node 22, 24, and 26 and includes build, type, formatting, lint, Docker startup, and MCP handshake checks, supporting strong static self-consistency. Hosted and stdio modes, OAuth/API-token access, payment modes, and multiple clients improve availability, but operation still depends on Apify services, Actors, search backends, networking, and credentials; the test source explicitly anticipates rate limits and unreachable external Actors. Concrete errors cover missing tokens, artifacts, or jq, timeouts, protocol failures, assertions, and named probes. A point is withheld because the evidence does not show a uniform end-user error contract across every production tool or the full troubleshooting guidance.

The material directly addresses hosted HTTPS, local stdio, desktop and web clients, editors, command-line use, scraping, search, storage access, run management, and several payment scenarios, with concrete examples. The tool inventory, default set, category selection, single-Actor minimum configuration, auto-injection rules, behavioral annotations, payment limitations, and absence of resources define boundaries thoroughly. Trigger precision benefits from Actor input-schema loading, the recommendation to fetch details before calling, and explicit tool selection. It is not full-strength because claims that the LLM automatically chooses the correct tool and parameters have no static guarantee, and dynamically loaded Actors retain third-party-defined behavior. Environment fit is strong: URL, CLI, environment-variable, OAuth, Bearer-token, transport, and client options are documented, alongside Node >=22 and a pinned pnpm toolchain.

The README has a clear contents structure spanning quickstart, clients, payments, tools, configuration, examples, troubleshooting, development, privacy, and contribution topics. Hosted connection, npx, local flags, environment variables, and MCPB installation are covered. Package, binary, MCP, and tool names are mostly stable, and v1-to-v2 compatibility plus retired aliases are explained; deductions reflect badges that still point at an older actors-mcp-server repository path and an explicit warning that default tools may change. Examples cover search, discovery, and pagination, but the supplied third example is truncated and the FAQ/troubleshooting body is not present. Limitations are unusually explicit: client support differences, payment scope, absent resources, removed SSE, temporary E2E instability, and test ownership are all distinguished. MIT metadata matches the complete license text. Version 0.14.4, migration notes, and a releases path are present, but no complete changelog or release policy is supplied. Author, copyright, issues, and contribution routes identify a maintenance path, while no named maintainers or response commitments are shown and publisher identity is unverified under the prompt.

Actor calls return run metadata, storage identifiers, a summary, and a nextStep that directs the client to paginated dataset retrieval, making outputs operationally useful. Schema-driven Actor tools, detailed discovery, and realistic examples further improve usability. The product offers clear marginal value through dynamic access to a large Actor catalog, generated schemas, run and storage management, documentation search, multiple transports, and accountless payment alternatives. Cost mechanics, free discovery, prepayment amounts, refund timing, and Pay Per Event constraints are disclosed. The score is reduced because Actor costs vary, current AGI terms are delegated to an external document, Skyfire requires a $5 minimum, and the supplied repository evidence shows neither mandatory pre-charge confirmation nor a general cost estimator or budget ceiling.

Core claims can be traced to the tool table, configuration contract, package scripts, CI jobs, and test-helper implementation. Full marks are withheld because numerical catalog claims such as thousands or 8000+ Actors, tested-client status, and hosted-only latest features are not individually substantiated by the supplied code or tests. Node support, authentication, transports, tool selection, telemetry test defaults, and test commands receive strong corroboration across README, package metadata, workflows, and test source. Fact-versus-inference separation is exemplary: the temporary E2E suite is repeatedly labeled AI-generated, not human-maintained, unsuitable as fact, excluded from CI, and subject to known instability. The README likewise distinguishes current defaults, possible future changes, removed behavior, and externally maintained terms.