Your context.
A visible boundary.

Two requests. Two decisions.

Generating a plan sends your goal to the Veilo Protocol server and its model gateway. Your source material stays in this browser tab at that stage. The browser sends only its character count and SHA-256 digest alongside your goal.

Approving the plan sends the source text to the server and gateway for analysis, together with the original goal and approved plan. Treat anything written in the goal as disclosed at the planning stage. Do not place credentials in either input.

What the model provider can see.

The gateway and its upstream model provider can process submitted text. The app requests Responses API processing with storage disabled (store: false). This setting does not verify the third-party gateway’s retention, logging, training, routing, or deletion practices.

This service does not provide end-to-end encryption against the provider, confidential computing, or zero-knowledge AI. HTTPS protects transport. Keeping text off-chain does not hide it from the service processing it.

Short-lived approval. Scoped metadata.

A random HttpOnly, SameSite=Strict session cookie binds each plan to the browser that created it. It expires after one hour. An approval expires after 15 minutes, is single-use, and is bound to the original goal, plan, model configuration and exact source digest.

Hosted deployments use Upstash Redis for pending goals, plans, source digests, credential hashes and usage limits. Expired records cannot authorize work; they are removed on access, and the shared store expires after one hour without activity. Declining removes the pending plan. Source text, reports and model credentials are never written to this store. Hosting access logs and provider records have their own retention rules.

Approval consumption and connection revocation are checked atomically across server instances. Local development without Redis uses temporary process memory; restarting clears that local state. This service has trial access keys and bounded usage, with no paid accounts or billing system.

Reports return to the requesting client.

In the web workspace, the report is streamed back to this browser. Reloading clears it. Downloading saves a Markdown file on your device. The application does not use localStorage, sessionStorage, third-party analytics or advertising trackers. Fonts and artwork are served locally.

The server’s report fingerprint is checked in the browser using SHA-256. It identifies report content; it is neither a proof of correctness nor a blockchain receipt. Stopping a request cannot recall text already submitted, and provider charges may still apply.

A trace without the source text.

The activity observatory records stage names, browser timestamps and reported token counts in this tab’s memory. It does not include your goal, source material, report text, credentials or raw error messages. It keeps up to 24 recent events; clearing the trace or reloading removes them.

Sample trace is explicitly fictional. This tab shows events from this browser interface, not a server audit log or network-wide telemetry. The animated protocol diagrams and proposed Solana record controls are illustrations and do not call models or grant permissions.

Files, masking and external connections.

The web workspace reads explicitly selected UTF-8 text files in this tab. Only selected source text is prepared. Common token, credential-field, private-key-block and email patterns can be masked locally. This is best-effort pattern matching, not a guarantee that every secret or personal detail is removed. Review the preview manually. Your goal is still sent at the planning stage.

The local MCP connector reads only named, supported text files inside the configured directory. Source preparation and keyword search run locally; prepared text remains in the connector process until approval. Tool results, including file names, masked snippets and reports, return to the AI host and may reach that host’s model provider. The remote redaction API/MCP tool receives its input on the Veilo Protocol server, although it does not call a model.

Connection keys expire in one hour, allow at most 20 model-call attempts, are stored only as hashes on the server and can be revoked from the issuing browser. At most three keys may be issued per browser session per hour; creating a new one revokes the earlier one. Existing server-wide capacity limits still apply. Keys are held in the connection page’s memory and can be copied or downloaded by you; save that configuration privately.

External plans use a separate connection principal. The private review link contains a secret in its URL fragment, and viewing or deciding requires the browser session that issued the key. The preview checks source text locally; only its matching digest is included with approval. The external client then sends the exact prepared text for a single analysis. Expiry, decline, replay and changed text are enforced by the server. Revocation blocks future requests; it cannot recall data already sent or stop all work already in flight.

Clear limits on actions and visibility.

The workspace performs text analysis and drafting. It cannot browse live sources, execute a repository, run code, publish, deploy, send messages, connect a wallet, move funds or submit Solana transactions.

The separate shared-view panel illustrates a future visibility policy using sample data bundled in the browser. Its switches do not enforce real access control or encryption.