mo
docs: enterprise documentation suite — compliance mappings (CRA, ISO 27001, IEC 62443), threat model, data flow, audit & deployment guides, SBOM, use cases, security.txt
5c5b617 |
Download docs/additional-documents/DATA-FLOW.md from aiagentmona/mona-agent: direct link, hf CLI and curl.
- Browser
- Download file 2.43 kB
-
https://huggingface.co/aiagentmona/mona-agent/resolve/main/docs/additional-documents/DATA-FLOW.md
- Command line
-
hf download hf://aiagentmona/mona-agent/docs/additional-documents/DATA-FLOW.md
-
curl -L -o DATA-FLOW.md https://huggingface.co/aiagentmona/mona-agent/resolve/main/docs/additional-documents/DATA-FLOW.md
2.43 kB
| # Data Flow & Data Minimization | |
| This document describes where data lives and moves in a mona-agent | |
| deployment, for GDPR/DPIA reviews and security assessments. | |
| ## Data inventory | |
| | Data | Created at | Stored at | Retention | | |
| |---|---|---|---| | |
| | User account | Cloud (Sngine) | Cloud DB | Until account deletion | | |
| | Device token | Cloud, on request | Cloud DB (plain, revocable) + device config | Until revoked | | |
| | Provider API keys | User input (dashboard) | Cloud DB, AES-256-GCM encrypted | Until deleted | | |
| | Chat messages | Dashboard / device | Cloud DB | Per-user; delete per agent or factory reset | | |
| | Run traces (reasoning, tool calls, results) | Device loop | Cloud DB (runs + steps) | Deleted with agent / factory reset | | |
| | Usage & cost metrics | Every LLM call | Cloud DB | Aggregated for insights; per-run detail deleted with agent | | |
| | Device telemetry (CPU/mem/disk/uptime) | Device, every 10 s | Cloud DB (latest + rolling history) | Forgotten on "Forget device" | | |
| | Device files & command output | Device tools | **Stays on the device** except task-relevant results streamed to the cloud for the brain | Ephemeral per run | | |
| ## Data flow (one task) | |
| 1. User sends a message (dashboard → cloud over TLS). | |
| 2. Cloud queues the task; the device polls and claims it (outbound HTTPS). | |
| 3. Device sends context to the cloud brain; the cloud calls the LLM provider | |
| using the user's key (decrypted in memory only). | |
| 4. Brain decides tool calls; the device executes them locally and returns | |
| results to the cloud. | |
| 5. The cloud stores the conversation, the run trace and usage metrics; the | |
| dashboard streams the answer to the user. | |
| ## Minimization principles | |
| - **No secrets on devices.** Provider keys never leave the cloud. | |
| - **No model weights anywhere in this product.** The product is a control | |
| plane; the models are third-party APIs. | |
| - **Telemetry is performance-only** (CPU, memory, disk, uptime, load). No | |
| keystrokes, no screen capture, no browsing history. | |
| - **Tool output is truncated** for transport (bounded context) — the full | |
| output stays on the device unless the task requires it. | |
| ## Data subject rights (operator toolkit) | |
| - **Export**: per-user training export (JSONL) and audit log access. | |
| - **Delete**: per-agent deletion, per-user factory reset, device | |
| telemetry forget, token revocation — all available in the dashboard. | |
| - **Minimize**: free plan limits bound data volume per account. | |