Draft — not legal advice; to be reviewed by a qualified lawyer before publishing.
Privacy policy
Version: 2026-10-31 (draft) Effective date: [EFFECTIVE DATE — set when published] Last updated: 31 October 2026 · All changes
1. Who handles your data
Castling by Cakraya is operated from Malaysia by Cakraya: [COMPANY LEGAL NAME], registration number [REGISTRATION NO.], address [ADDRESS]. Cakraya's exact registered form and SSM registration are pending. Privacy and data-protection officer contact: [CONTACT EMAIL] (designated officer details must be confirmed before publication).
This notice covers our website, waitlist and the Castling services we operate. It is organised around Malaysia's Personal Data Protection Act 2010 (PDPA), as amended. It describes current beta behaviour; it is not a claim of certification or that planned privacy features already exist.
For account administration, security and our waitlist, Cakraya determines the purposes of processing. A workspace customer normally determines why its contacts' messages are processed, and Cakraya processes that content to provide the customer's requested service. The precise controller/processor responsibilities depend on the deployment and applicable law; business processing arrangements require appropriate contractual terms. A contact need not have a Castling account to have data processed here.
For a deployment hosted by someone else, that host controls its server and must provide its own notice. Installing the runner on your computer does not make a Cakraya-hosted workspace self-hosted: its server still receives and stores workspace data.
2. Notice and choice: information we collect
We collect information from you, your sign-in provider, your workspace members and connected accounts, your contacts' messages and platform events, and your use of Castling.
| Information | What this can include |
|---|---|
| Account and workspace | Name, email, Google identity and profile image, membership, role, profile preferences, invitations, sessions and connected-app grants. |
| Waitlist | Email and any name, company and use-case text you submit; referral URL/campaign information, signup time and a keyed hash of your IP address for abuse prevention. Signup name, company and email can also appear in the operator workspace's Team chat notice; members with settings-edit permission can read/export the signup list. |
| WhatsApp and CRM | Linked number and device connection information; contact and group identifiers, names, phone numbers, profile photos, messages and imported history; timestamps, delivery/read information, quotes, attachments, images, videos, stickers and voice notes; opt-in/opt-out records, labels, notes, pipeline stages, priorities and feedback. This includes your correspondents' information. |
| Other channels | Connected social-account identifiers and tokens, Messenger/Instagram messages and Facebook/Instagram comments, social-post drafts, media, publication results and statistics where enabled; Telegram chat/member identifiers, messages and files where bridged. TikTok's current integration supports connection and publishing, not a TikTok direct-message inbox. |
| AI and project work | Prompts, selected chat context, agent instructions, outputs, approvals, summaries, learned writing-style guides, transcripts, relevant project files and task/run records. These may contain personal information about you and others. |
| Optional place and weather | With the feature enabled, a typed city or browser-provided coordinates and accuracy. The browser may send precise coordinates to the Castling server; the server rounds coordinates to two decimal places (roughly 1 km) before storage and weather/place lookups. AI context uses an area name and weather, not coordinates. |
| Technical and usage data | IP addresses processed for connections and abuse controls, request/error and diagnostic information, runner and device status, usage/cost counts, timestamps and security events. Logs may contain identifiers, URLs and platform-event metadata; do not assume logs never contain personal data. |
Some data is necessary to create an account, operate a selected integration or keep the service secure. Without it, that function may not work. Linking messaging accounts, importing history, using AI, optional location and bridging chats are choices; review their settings and recipients before enabling them. Imported history can be stored even when excluded from ordinary reply prompts. Whole-chat context and selected style-learning options can make historical content available to AI.
You may disable an optional feature, revoke an external connection or ask us to restrict processing. Disabling future processing is not the same as deleting data already stored or already sent. Withdrawal of consent does not undo earlier lawful processing. You should give your contacts an appropriate notice and obtain consent or another valid legal basis where required; agreement to these Terms is not their consent.
3. General principle: purposes and lawful grounds
We use personal information to provide requested workspaces and connections; receive, organise and send messages; produce suggestions, replies, transcripts, style guides and summaries; run authorised agent tasks; handle support and rights requests; prevent abuse; investigate failures; and meet legal obligations. Waitlist information is used to respond to your signup and contact you about beta access and launch, and referral information helps us understand signup sources. Ask [CONTACT EMAIL] to leave the waitlist or stop launch messages. Joining it is not consent to unrelated advertising.
Where the PDPA applies, processing requires consent or an applicable statutory basis or exception, with additional requirements for sensitive personal data. Do not upload sensitive information without the authority and safeguards it requires. We do not treat a GDPR ground such as legitimate interests as a universal exception to Malaysian consent requirements.
Where the GDPR applies, grounds may include performance of our contract with you for necessary account/service operations; consent for optional processing where required; legitimate interests in proportionate security, support and service reliability, balanced against individuals' rights; and compliance with legal obligations. Contract with a workspace owner does not itself justify all processing of that owner's contacts. The customer must establish its own valid basis and, for special-category data, an additional permitted condition; Cakraya needs a basis for any separate purposes it determines. No new unrelated purpose is authorised by this draft.
AI triage may infer urgency, intent, language, CRM stage or a need for a person. Style learning estimates how you write. Such inferences may be inaccurate. Users must provide human review before using them for decisions with legal or similarly significant effects on people.
4. Disclosure: who receives information
Information can be disclosed to these recipients as needed for enabled features:
- Your workspace members and administrators under current access controls; authorised agents, runner operators and connected apps; and recipients of messages, exports, webhooks or Telegram bridges you enable. The current beta does not yet guarantee complete per-person isolation of every number and its derived data. Workspace-level access and operator access must be considered before connecting private communications.
- Your selected AI providers, through your CLIs/accounts or supported API connections. Depending on your setup these can include Anthropic (Claude), OpenAI (Codex), Google (Gemini/Antigravity), DeepSeek, a routing provider, or a local model. Relevant messages, instructions, context and supported images/files may be sent. Provider retention, training controls and terms depend on your account and settings; a CLI on your PC may still send data to its cloud provider.
- TypeSafe/Jev, where enabled for triage, classification or other judgments, which receives the relevant question and context and any supplied images. Choosing a local reply model does not automatically disable separate cloud judgments or other online tools.
- WhatsApp/Meta, Telegram and TikTok for connected messaging, login, publishing and platform functions; Google for sign-in and GitHub for connected projects where used. An unofficial WhatsApp linked-device connection is distinct from an official Cloud API connection.
- Open-Meteo for weather/city lookup and OpenStreetMap Nominatim for area lookup when location is enabled. Rounded coordinates or the typed city are sent to the relevant service; an area and weather description can enter reply context.
- Hosting, network and operational service providers used in the deployment, and personnel who need access for operating, securing or supporting it. The final hosting suppliers and processing locations have not yet been confirmed for public launch; request details at [CONTACT EMAIL] before uploading data that needs a particular location or supplier restriction.
- Authorities or other recipients where required by law, or where a lawful, necessary disclosure is needed to protect rights or investigate abuse. A business transfer involving personal data would require appropriate safeguards and notice.
These categories include service providers acting on our instructions and platforms/providers with their own responsibilities. They are not a final named sub-processor register or a promise that every provider uses identical contractual terms. We do not sell personal data or use advertising trackers in the current app. Marketing referral information submitted with the waitlist is still collected as described above.
5. AI safeguards and local processing
The secret guard replaces detected passwords, keys and similar secrets in supported outgoing AI text with numbered placeholders. It applies to supported agent prompts/completions and Jev text, but detection is imperfect; binary files, images and audio are not fully inspected by that text filter. It does not remove the original from Castling's stored messages or make all personal information anonymous. Do not rely on it to authorise disclosure of confidential material.
Local-model inference, when actually configured to run locally, does not send that inference request to a cloud AI provider. It still receives the input locally and may be on the Castling server or a runner rather than your own device. Hosted storage, messaging platforms, optional Jev, weather services and agents' network tools are separate data flows. There is not yet a complete enforced per-number "nothing leaves my device" mode.
Voice-note transcription currently runs with local Whisper on the machine running the Castling server, with temporary audio removed after processing. Original audio and transcripts can remain in message storage, and transcripts may enter subsequent AI context. Style learning can use local Ollama or a configured agent; the latter may use a cloud provider. A future local-only feature must not be assumed to cover these separate settings today.
6. Cross-border processing
Cakraya is based in Malaysia. Hosting, selected providers, platforms and your runners may process information in other countries. Data is not guaranteed to stay in Malaysia, or on your computer, merely because Cakraya is Malaysian or you use your own AI account.
Applicable Malaysian transfer requirements and, where the GDPR applies, its international-transfer rules must be met. Before public business deployment, the relevant locations, provider contracts and lawful transfer arrangements must be confirmed. This draft does not assert that standard contractual clauses, an adequacy decision or any other transfer mechanism is already in place for every recipient. Contact [CONTACT EMAIL] before use if your information requires those arrangements; optional connection consent does not remove mandatory transfer protections.
7. Security principle
Current safeguards include hosted sign-in and session checks, CSRF protection, workspace/role and chat-access checks, protected connection flows, restrictive permissions for sensitive server files and encryption for certain stored integration tokens. Secure cookies are used in hosted production configuration; session-cookie access by scripts is restricted, while the CSRF cookie is deliberately readable by the app. Proper HTTPS deployment is required for hosted transport protection.
These measures do not mean every database row, media file, log or backup is encrypted. Host operators can access server storage, and your chosen runner and AI provider can access data supplied to them. WhatsApp transport encryption does not make Castling's imported copies inaccessible to its host. We cannot guarantee absolute security. Report suspected exposure to [CONTACT EMAIL]; we will investigate and make notifications required by applicable law. Detailed per-number privacy and support-access auditing remain planned work, not current guarantees.
8. Retention principle: current deletion limits
Data should be kept only for a necessary, lawful purpose. The beta currently lacks a complete automatic retention and erasure system. There is no general enforced expiry for chats, media, contacts, learned information, waitlist entries, operational logs or backups; these may persist until separately removed. Contact [CONTACT EMAIL] for a retention or deletion request rather than assuming an account or number action erases everything.
| Action or record | Current behaviour and limits |
|---|---|
| Remove a WhatsApp number | Stops/removes its connection, deletes local linked-device login files and connection-specific records, including the separate learning-history store and broadcast/survey records. Imported workspace conversations, message content, attachments, CRM contacts and some derived data remain. Message identifiers and related metadata are preserved against the phone number for deduplication and can be adopted on relinking. This is not full erasure. |
| Delete a message | Clears its body, marks it deleted and removes attachment records/files on a best-effort basis. A message row and metadata can remain. Dismissing a draft hides it rather than erasing it. Other stored copies and derived information are not comprehensively purged. |
| Clear a chat or delete an eligible user-created group | Removes that chat's message, attachment, request, run and request-context records and attempts to remove its attachment files; deleting the group also removes its conversation/membership records. This does not comprehensively erase contact profiles, style/history stores, summaries elsewhere, logs, backups or external copies. The group-delete operation is not a general delete-number or delete-workspace feature. |
| Remove a member or leave a workspace | Removes membership/participation and adjusts or removes affected sessions. The global account, messages already posted, workspace content and connected-number data are not automatically erased by that action. |
| Close an account or erase a workspace | There is no complete self-service workflow or guaranteed automatic cascade today. Requests require manual review and handling, with attention to other people's data and lawful retention obligations. |
| Security sessions and guard counts | Hosted sessions normally expire after 30 days of inactivity, refreshed with use; expired server rows are purged on subsequent session reads, at most hourly. Secret-guard records contain counts/kinds, target and chat references, not secret values; rows older than 90 days are purged when another counted guard event is recorded for that workspace, rather than by a guaranteed daily expiry job. |
| Backups and operational logs | The backup tool copies the database, token-key files and linked-device login state into local backup directories. It does not enforce an expiry or automatic removal of deleted content. Operational logs, including the server's cloud log, also have no application-enforced retention period. Restored backups may contain previously removed data; automated reapplication of erasure requests is planned, not implemented. |
| Optional location | Turning location off clears the stored active place. It does not retroactively remove earlier messages, AI/provider context, logs, backups or all temporary lookup caches. |
Deletion in Castling does not delete messages from recipients' devices, WhatsApp/Meta, Telegram or other platforms, files already downloaded, runner/CLI histories or AI-provider records. Their separate policies and deletion tools apply. Where legally required and applicable, requests must also be followed through with relevant recipients; instant erasure of all external copies is not promised.
Planned work includes distinct disconnect and erase actions, scoped export, fuller removal of derived data, a documented backup expiry and an erasure ledger applied after restoration. Those controls, including suggested 30-day backup and 90-day audit periods in the privacy design, are proposals rather than current retention guarantees.
9. Data integrity principle
Please keep your profile and contact information accurate, correct CRM notes and review generated transcripts, summaries and labels before relying on them. Imported names and AI inferences may be incomplete or outdated. Contact [CONTACT EMAIL] for corrections that the interface cannot make. Corrections to Castling do not automatically update a platform, a previously sent message or an AI provider's copy.
10. Access, choice and individual rights
You, including a contact whose messages were imported, may contact [CONTACT EMAIL] to ask what personal information is held about you, request access or correction, withdraw consent where relevant, or raise a concern about processing, disclosure or retention. Applicable law may also allow you to request portability, deletion or limits on processing and to object to direct marketing or other uses. Availability and exceptions depend on the law and the circumstances.
We may ask for proportionate identity/authority checks and enough context to locate the records, and will protect other people's information when responding. Do not send passwords or unnecessary identity documents in an initial request. A contact request should not result in disclosure of the workspace owner's entire inbox. Where a workspace customer controls the information, we may coordinate with it or direct you to it while handling our own legal duties. The absence of a self-service button does not remove statutory rights; requests currently need manual handling, and we will explain any lawful restriction, retention reason or available complaint route.
Where the GDPR applies, rights can include access, rectification, erasure, restriction, portability, objection, consent withdrawal, and safeguards concerning solely automated decisions with legal or similarly significant effects. Direct-marketing objections must be respected. We will respond within the applicable legal deadline, normally one month for GDPR requests, and explain any legally permitted extension. You may complain to your relevant EU supervisory authority; Malaysian concerns may be raised with the Personal Data Protection Commissioner. GDPR applicability is assessed on the processing, not simply whether someone holds an EU passport. These rights are not a claim that Castling's planned automated export/erasure tools are already available.
11. Children
Castling accounts are intended for adults aged 18 or over, not children. Contacts' messages may nevertheless include children's information. Users must have any required parental authority or other lawful basis and avoid unnecessary children's data. Tell [CONTACT EMAIL] if a child's information has been processed improperly so that it can be investigated and an appropriate restriction or removal request handled.
12. Changes and contact
We will update this notice when data flows or practices change, show a new effective date and give appropriate notice of material changes. We will obtain fresh consent where required rather than treat publication alone as consent to a new use.
For privacy, access, correction, deletion, waitlist withdrawal or the designated data-protection contact, write to [CONTACT EMAIL] or Cakraya ([COMPANY LEGAL NAME]), [ADDRESS].