CopilotKit

PARTIAL consensus from 1 PASS, 2 PARTIAL.

The quickstart completes every requested step, but it uses a legacy endpoint helper while the runtime page tells new projects to use v2 handlers. The nine steps are complete and copy-pasteable, but the runtime code sets model: "openai:gpt-5.4-mini" while the same page's callout says the example uses GPT-4o and Model Selection lists neither gpt-5.4-mini nor any gpt-5.4 specifier as supported, and step 1 tells you to get a license key that no later step ever uses. Quickstart provides a complete linear path from create-next-app to a working CopilotSidebar with Copilot Runtime and BuiltInAgent, with no gaps or ambiguity.

FAIL consensus from 1 PARTIAL, 2 FAIL.

The page names all five limit categories but omits a complete per-plan matrix, and the linked dashboard returned only an app shell. The documentation names all five limit categories but publishes no value for any of them and points only to a login-gated dashboard pricing page; the real numbers (Developer 3 days / 200 threads / 1 GB, Pro 5 days / 5,000 / 10 GB, Team 14 days / 25,000 / 100 GB at up to 5 seats, Enterprise custom) live on the unlinked marketing page www.copilotkit.ai/pricing, which also lists a Pro plan the docs never mention. All numeric limit values are available on copilotkit.ai/pricing but not on docs.copilotkit.ai, the documentation site has no pricing or limits page (both /pricing and /limits return 404), and the Pro plan does not state a seat count, requiring inference.

PASS consensus from 3 PASS.

The reference identifies the discovery cause, shows default registration, requires ID matching, and directs an /info probe. The error reference gives the exact error name, all three message variants, the cause, the corrected new CopilotRuntime({ agents: { default: myAgent } }) registration, and a GET {runtimeUrl}/info discovery probe, and cross-links the server-side 404 counterpart and the fresh-thread /connect case. The error-reference page documents the exact error name, three message variants, root cause, and a three-step fix covering agent registration, id verification, and runtime-info probing.

PASS consensus from 3 PASS.

One guide documents the headers prop, onRequest token validation, identifyUser, and ownership checks for unscoped routes. The page covers the whole path end to end, forwarding the token via the headers prop, verifying it in onRequest with createCopilotRuntimeHandler, naming the caller with identifyUser, and a per-route table that flags threads/events, threads/state, and agent/stop as not user-scoped and shows the onBeforeHandler guard for them. Auth page documents the complete flow: frontend forward via headers prop, backend validation via onRequest hook, and per-user scope via identifyUser with full code examples and a security checklist.

PARTIAL consensus from 1 PASS, 2 PARTIAL.

The reference supplies a handler-plus-render example, but it omits z and ToolCallStatus imports and relies on separate pages for package installation. The reference example matches the quickstart's @copilotkit/react-core/v2 import surface and fully specifies handler and render, but the guide page /frontend-tools shows no renderer at all, and the renderer example uses ToolCallStatus without importing it, the import lives on a separate page as import { ToolCallStatus } from "@copilotkit/core", a package no install command in the docs mentions. The v2 React SDK installation is consistent across quickstart and reference pages, useFrontendTool documents both handler-only and handler-with-renderer patterns with full Zod schema examples, and the related useComponent hook provides an alternative renderer-first path.