Hosted chat
Use whenA user needs a multi-turn conversation, follow-up questions, or visible answer history.
VerifyIdentity, session continuity, citations, attachments, feedback, reset, and escalation.
Interface launch · Updated August 25, 2026
A working agent can still fail at access, files, sessions, embeds, error states, or release. Choose the surface by the user's job, then test one complete path from first visit to rollback.
Commercial disclosure: if our referral link is active and you later buy an eligible paid Falcon Builder plan through it, we may earn a commission. This launch checklist is independently useful without that link.
Decision one
Do not begin with a feature list. Begin with the interaction the user must complete and the evidence that proves it worked.
Use whenA user needs a multi-turn conversation, follow-up questions, or visible answer history.
VerifyIdentity, session continuity, citations, attachments, feedback, reset, and escalation.
Use whenThe task needs known fields, repeatable validation, and one bounded submission.
VerifyRequired fields, file rules, confirmation, duplicate submit behavior, and result delivery.
Use whenThe experience must live inside an existing product, help center, or marketing site.
VerifyAllowed origins, frame policy, responsive layout, host-page events, and failure isolation.
Use whenThe user already works in SMS or WhatsApp and expects an ongoing thread.
VerifyProvider dependency, contact identity, media handling, opt-out, delivery failure, and handoff.
Decision two
Anyone with the URL can open it. Use only when the data and action are safe for anonymous traffic.
A single secret creates a lightweight gate, but does not identify which person used it.
An address proves access to an inbox and can bind a conversation to one permitted identity.
Existing organization membership controls access and should inherit current removal rules.
Central identity and lifecycle control for environments that have verified plan and setup support.
Input contract
“Supports documents” is not a testable requirement. Type, size, validation, storage, and processing must agree at every boundary.
List each accepted text, image, PDF, or document format. Reject everything else.
State limits before upload and test both the browser and processing boundary.
Check extension, MIME type, structure, required fields, and empty or corrupt content.
Define scanning, quarantine, rejection, and what reaches downstream tools.
State where input, extracted text, and model context remain—and for how long.
Verify that the chosen model and workflow can actually process each allowed input.
Session contract
Identity, memory, and history are separate choices. Test anonymous and authenticated paths independently.
Choose the stable key that joins turns without merging two people.
Bind identity at the interface, not from a name typed into a prompt.
Decide whether a verified user resumes the same conversation elsewhere.
Give the user and operator an explicit way to start over or remove stored history.
Failure UX
A user needs an honest state, a next action, and protection from duplicate effects—not a generic “something went wrong.”
Hold the downstream response past the interface limit. Show a useful state and a safe retry.
Interrupt a streamed answer. Mark it incomplete instead of presenting it as final.
Submit an unsupported, empty, renamed, or corrupt file and explain what was rejected.
Expire or clear the session between turns. Do not silently attach the user to another thread.
Disable the model or integration and preserve the user input without inventing success.
Trigger the handoff path and verify the person receives the required context.
Release control
Change the workflow and interface without touching the live user path.
Run representative users, devices, files, failures, and access states against the draft.
Promote one reviewed version and record what changed, who approved it, and when.
Restore a known version and confirm sessions, embeds, and integrations recover as expected.
One complete-path launch test
Use harmless data and one representative task. Keep the acceptance evidence together.
Start with a new user. Enter through the real URL and complete the intended access step.
Submit the hardest valid input. Include the largest allowed file and all required fields.
Continue the session. Ask a follow-up that depends on prior context.
Repeat on a second device. Confirm the expected identity and history behavior.
Force one failure. Recover without duplicate processing or a false success state.
Publish and roll back. Prove the tested version can go live and a known version can return.
Falcon-specific checkpoint
Falcon Builder's current homepage describes hosted chat, forms, a website chatbot, and WhatsApp/SMS. It also labels public, password, email + OTP, workspace login, and SSO/OIDC as access modes, and describes image, PDF, and document attachments. Verify the exact interface, input limit, and plan in the live product before committing.
The current MCP documentation describes password-protected or email-allowlist interfaces, per-workspace identity, embed-origin restrictions, draft editing, runs that may create real side effects, and separate publish/rollback permission. Those details are useful evaluation targets; they do not prove that every interface, plan, file, device, or integration behaves identically.
Current pricing marks some enterprise controls as “Coming soon.” Falcon's Terms do not promise uninterrupted, secure, error-free, accurate, or reliable service. Keep launch acceptance criteria specific to the chosen configuration.
Primary sources