Reference
Known Limitations
The SDK is now a strong MVP, but the product boundary should still be documented honestly.
Need the shortest path?
Use Quick Start when you want the smallest working route → scan → graph → run flow first.
Something feels off?
Inspect scan output, graph state, active route/context, and dispatch in a practical order.
Need the full surface?
Browse all runtime methods, hooks, extension points, and lifecycle APIs.
Passing focused validation and red-team coverage means the SDK is real and credible for its intended scope. It does not mean it is a universal web agent or a finished platform for arbitrary pages.
The current implementation is much stronger than the earliest MVP, but there are still intentional limits.
It is for websites you control
Cinder is strongest when developers can:
- add semantic tags deliberately
- understand the app surface
- choose when extensions are needed
- improve route/context clues instead of guessing blindly
It is not optimized for arbitrary untagged websites.
Local resolution is focused, not universal
The planner system is practical and graph-aware, but it is not trying to be a broad general natural-language engine.
That focus is a strength, not a weakness.
Custom widgets may still need adapters
Standard browser controls work much better now, but apps with heavily custom widgets may still need registerDomAdapter().
Fallback is bounded, not magical
Fallback is more mature now, but it still depends on:
- a configured endpoint
- provider behavior
- graph grounding
- ambiguity-safe normalization
A fallback result can still be rejected if it cannot be safely grounded to the real graph.
Browser speech still depends on platform support
Speech input prefers the Web Speech API when it's available. For browsers with no Web Speech API at all (Firefox, older Safari), setting speech.engine: "auto" with a configured fallback endpoint enables a server-side recording-and-transcription path instead - see Speech Input. It cannot stream interim results the way the Web Speech API can: there's a transcript only once recording stops and the upload finishes, not a running partial transcript while the user is still speaking. That's an honest UX difference, not a bug to chase. Cinder Relay does not implement speech transcription today - there is currently no publicly available server offering this path.
Tier 2 (embeddings) covers routes, outputs, and actions - not fields or filters, yet
The optional @usecinder/embeddings package recovers typos and paraphrases of a page's route/output/action nodes before any network call, plus a separate, simpler grammar-shape match (and, as a last resort, a small extractive question-answering model) for set_field/set_filter commands specifically. It never claims the top resolved-exact confidence level - every tier-2 match is reported honestly as a best guess among known candidates, never full certainty. See Tier 2: Embeddings.
Relay is a real, working prototype of hosted billing - not a finished managed service
@cinder/relay has a real, persistent Postgres-backed store (verified against an actual live database, not a mock) when you run it with DATABASE_URL set, real token-based cost metering, real account authentication, and a working dashboard UI. Two things are still genuinely unfinished:
- Razorpay payments are unverified against a real account. The order-creation and webhook-signature-verification logic is real and tested (real HMAC-SHA256 computations, not faked), but no real test-mode Razorpay account was available while building it.
balance_centsis stored as a floating-point number, matching the simplest possible implementation rather than the integer micro-cents a production ledger handling real money at scale should use.
See Relay Overview for the full, current list before putting real money behind it.
The docs and DX are still catching up to the runtime
The SDK has matured quickly. The documentation and onboarding experience are improving, but developer ergonomics are still an active part of the product work.
Continue reading