01 Who We Are & Scope of This Policy
This Privacy Policy describes how Mesh Labs Cloud, Inc., a Delaware corporation having its mailing address at 2810 North Church Street, Wilmington, Delaware 19802, United States ("Mesh", "we", "us" or "our"), processes personal data in connection with MESH and any related websites, applications (including the MESH desktop application, the MESH mobile application, the MESH Agent Browser browser extension, and the MESH IDE extensions for Visual Studio Code, Cursor, Antigravity and compatible editors), SDKs, browser-based grid nodes, control plane and APIs (collectively, the "Service").
For purposes of the EU and UK General Data Protection Regulations ("GDPR"), the Israeli Privacy Protection Law, 5741-1981 and its regulations ("PPL"), the California Consumer Privacy Act as amended by the CPRA ("CCPA/CPRA") and similar laws, Mesh is the controller (or "business") of personal data we collect directly from you and technical telemetry generated by your use of the Service. Where the Service is embedded or administered by a third-party customer, Mesh may act as a processor (or "service provider") on that customer's behalf, as described in the applicable customer agreement or data processing addendum.
In this Policy, "personal data" includes information that identifies, relates to, describes, is reasonably capable of being associated with, or could reasonably be linked, directly or indirectly, with an identified or identifiable individual, as defined by applicable law.
You are generally not legally required to provide personal data to us, and providing it is voluntary. However, without information required for an account, device connection, payment, security or a requested feature, we may be unable to provide some or all of the Service.
This Policy supplements, and is incorporated into, our Terms & Conditions. Capitalized terms not defined here have the meanings given in those Terms.
02 Plain-Language Summary
- We collect account data (e.g., email), technical telemetry (e.g., IP, peer ID, GPU/backend capabilities, performance metrics), usage data (jobs you submit, models you use, contributions you make as a Provider) and, where you choose, payment data (handled by our payment processor).
- We use that data to operate the Service, match jobs to nodes, calculate Compute Units (CU), validate results, prevent fraud and abuse, comply with law, and improve the Service.
- The Service is decentralized: your inputs and intermediate results may be processed on third-party browser nodes that we do not control. You should treat any data you submit as visible to such nodes for the duration of the job.
- Personal AI conversations are kept on your MESH desktop applications. Your signed-in desktops copy them to each other directly over encrypted peer-to-peer connections, and your phone (and, if you approve it, a web browser) reads them from a desktop. We do not keep a conversation store on our servers. However, when a conversation is not Private, each turn that runs as a job is stored with your job History, and that job includes the conversation history the turn sent to the model (§11).
- Where a request runs. A chat, agent or generation request runs either on the device you are using, or as a job that our control plane routes to a cluster: by default your personal cluster (your own devices and any devices you invite to it), or another cluster or the open Global Network if you choose one. A job sent to a cluster passes through our servers on its way.
- Private (on by default in the desktop, mobile and web apps) means we do not store the inputs or outputs of that job in our database or show them in History; they are held only briefly in memory while the job runs. Anonymous statistics such as prompt length and timing are still kept. Private controls what is stored, not where the job runs.
- Our website uses Google Tag Manager and the Google Analytics tag it loads to measure visits, only if you allow analytics in the cookie banner; you can change that at any time from Cookie settings (§07). Either way, the website counts which of the two choices you made, anonymously (§07).
- If you enable Agentic Features, the desktop application may access resources you authorize and act through connected accounts, including changing or deleting files, executing commands, publishing or uploading content, sending messages or emails, and invoking third-party tools. Data is sent to a third party when the requested action requires interacting with that party.
- Global-network jobs may be processed by unrelated Provider nodes. Organization Cluster jobs are scheduled to participating devices in the applicable cluster, which may include devices operated by invited members or suppliers. A cluster may be configured to fall back to the Global Network, in which case unrelated Provider nodes can receive the work. Mesh-controlled cloud infrastructure may still process metadata, inputs and outputs.
- We do not sell your personal data. We share data only with the sub-processors and partners listed below, with peer nodes assigned to your jobs, and where required by law.
- We are based in the United States. The Service connects users and Provider nodes globally, and your data may be processed in countries other than your own. Where required, we use contractual and other safeguards for international transfers.
- You have rights of access, rectification, erasure, restriction, portability and objection, subject to applicable law. See §13.
03 Categories of Personal Data We Process
| Category | Examples |
|---|---|
| Account & identity | Email address, display name, password hash, organization name (for business accounts), the identifier and verified email your sign-in provider returns when you use single sign-on (for example Google), authentication tokens (JWT) and session cookies, API keys, role and permission metadata, records of the model-specific terms you accepted (which terms, which version, when, and from which account), and records of which version of these Terms, this Policy and the in-app disclosures (Gmail, the browser extension) you accepted or declined, when, and in which application. |
| Device & browser capabilities | User-agent, browser type/version, operating system, screen size, supported web APIs, WebGPU/WebNN/WASM availability, GPU adapter name and feature flags, available memory, CPU class, device language and time zone; and, where you install the desktop application: your device name (hostname, shown in your dashboard to distinguish your devices), system resource state (total and available memory, CPU load, disk capacity), power source and battery state, thermal state, and screen-lock / user-idle state (used to schedule work when your device is unattended and to yield resources when you are active). |
| Network identifiers | IP address, approximate geolocation derived from IP, a pseudonymized network-locality identifier (a salted hash derived from your IP address; the raw address is not stored in this identifier), peer round-trip-time measurements between your node and peers it connects to, peer identifier (PeerJS / WebRTC ID), session identifier, signaling room membership. |
| Hardware attestation | We do not currently collect hardware or platform device attestations. If we introduce them (for example WebAuthn, TPM or Secure Enclave attestations, to resist Sybil attacks), they will be opt-in where the law requires it, we will process only the resulting public key and assertion, and we will update this Policy first. |
| Usage & job metadata | Job IDs, model identifiers, job type (text-to-image, image-to-video, etc.), parameters (steps, guidance, resolution), timing, timeouts, success/failure status, validation outcomes and quorum results. |
| Job content (Requesters) | Prompts, source images, source videos, source 3D inputs, source code, file contents and repository context submitted through the IDE extensions, intermediate tensors and generated outputs to the extent they pass through the control plane or persistent storage. Inputs submitted to the Global network may also be transmitted directly to unrelated Provider nodes. Organization Cluster workloads are scheduled to eligible devices participating in that cluster, which may include devices operated by invited members or suppliers. A cluster may be configured to fall back to the Global Network, in which case unrelated Provider nodes can receive the work. Mesh's hosted control plane and S3-compatible object storage may still process Organization Cluster metadata, inputs, intermediate artifacts and outputs. Where a model runs on the device you are using (a local model in the IDE extensions, the desktop application's local runtime, or the desktop's offline mode), chat and job content is processed on that device and is not transmitted to us or to peer nodes. Where a Personal AI request runs as a job on a cluster, including your personal cluster made of your own devices, the job passes through Mesh's control plane, which processes its content as necessary to route it to the devices that run it; Mesh shall not transmit such content to third-party AI models. For a Private job, the control plane keeps the inputs only in temporary memory while the job is in flight (at most 2 hours) and the outputs only for up to 15 minutes so the waiting application can collect them, and stores neither in its database; files the job uploads or produces stay in object storage only until the automatic deletion described in §11 removes them, within 3 days; it keeps anonymous statistics (for example prompt length, number of turns, output size and type) and non-content parameters (for example image size or step count) to operate, price and size the Service. |
| Artifact access links | Upload and download URLs for job inputs, checkpoints and generated outputs. Some are time-limited presigned bearer URLs: they do not require a separate Mesh login after issuance, and anyone who obtains the complete URL may access the associated object until the URL expires. This can apply to Global-network, personal-cluster and Organization Cluster jobs. |
| Personal AI conversations | Conversations, messages, attachments and generated media in Personal AI. The desktop application stores them on your computer. Your signed-in desktops copy conversations, long-term memory records and generated media to each other directly over encrypted WebRTC data channels; our servers only help the devices find each other (signaling) and do not receive the content. The mobile application reads conversations from a linked desktop over the same kind of channel and keeps a cached list of conversation titles and copies of the media you open on the phone. If you turn on browser access in the desktop's settings and approve a specific browser, that browser can open your Private conversations from the desktop and holds them only in memory for the open page. When a conversation is not Private, the jobs its turns submit are stored with your job History on our servers and can be viewed in the web application. |
| Agentic content, permissions & actions | If you enable Agentic Features: selected files and folders, file contents and metadata, directory listings, page and application content, command and tool inputs and outputs, MCP responses, installed skills and MCP servers, granted roots, capability settings, consent requests and decisions, schedules, Agent Actions and their results, and agent transcripts and audit history. Agent Actions may include reading, creating, editing, moving, overwriting or deleting files; operating software or a browser; submitting forms; publishing or uploading content; sending messages or emails; invoking APIs; and, where enabled, using paid services or making purchases. |
| Connected accounts & communications | Connected-service account identifiers, authorization scopes, OAuth or API access tokens, browser-session data and cookies, recipient and contact information, message and email content and attachments, public posts and uploads, transaction details and amounts, and responses received from connected services. Password fields and credentials entered during a human-takeover flow are not intended to be exposed to the AI model, although the relevant third-party service necessarily receives the credentials you enter into it. Connector tokens (for example for Gmail, Linear, Notion, Atlassian or GitHub) are stored on your computer, sealed with your operating system's keychain or equivalent, and are not sent to our servers. See §08 for Gmail and other Google user data. |
| Memory, schedules & personal space | If you turn on long-term memory in the desktop application: facts, preferences and short summaries of past turns kept in a memory store on your computer and copied only to your other signed-in desktops (memory from a Private conversation stays on the desktop that ran it). If you ask the assistant to schedule a task: the schedule, its prompt and its run history, kept on the desktop that created it. Where the personal space feature is enabled for your account: a bounded profile (about you, your preferences, your work, facts and notes) that you write or approve, built from summaries of your non-Private jobs and stored on our servers so your devices can read it. |
| Browser extension | If you install and pair the MESH Agent Browser extension: the addresses of the pages the agent opens in a tab the extension created, and the text and controls the agent reads on those pages, passed to the MESH desktop application on the same computer; and a pairing token stored in the extension. See §08. |
| Mobile application | Sign-in tokens in the phone's keychain, the push-notification token (Apple Push Notification service or Firebase Cloud Messaging) registered to your account so the desktop can ask you to approve an agent action or tell you that a task has ended or is waiting for you, a cached list of conversation titles and model names, and copies of generated media you open (up to 2 GiB, in storage excluded from device backups). |
| Contribution metrics (Providers) | Compute time, accelerator and CPU operations executed, validated job completions, shard seeding contributions, uptime, latency to peers, idle/busy state. |
| Reputation & Compute Units (CU) | Reputation scores, CU balances, CU transactions, deterministic complexity metrics and oracle-based reward calibration outputs. |
| Communications & support | Emails, support tickets, in-product feedback, security reports. |
| Payment data (where applicable) | Billing name, address, VAT/tax identifier, transaction amounts, credit balances and ledger entries. Payment-card data is processed directly by our payment processor (Stripe) and is not stored by Mesh. For Provider payouts: the Stripe connected-account identifier and payout status, and, if you choose a stablecoin payout, the wallet address you provide. Stablecoin transfers are recorded on a public blockchain, where the wallet address and amount can be seen by anyone and cannot be deleted. |
| Identity verification (where applicable) | Where we or our providers conduct Provider or payout verification (including know-your-customer checks), identity documents, government identifiers, sanctions-screening results and related verification outcomes may be processed as described in this Policy. |
| Logs & security data | Request and response logs, error traces, heartbeat events, signaling messages, abuse-detection signals, IP reputation data, rate-limit telemetry, CAPTCHA results (Cloudflare Turnstile, where enabled), node runtime events and degradation reports (which fallback a node used and why), device benchmark results (compute and memory figures per device), and audit records of security-relevant actions. Test (QA) builds of the desktop application also upload a redacted extract of their own logs for diagnosis; production builds do so only if that setting is turned on. |
| Website analytics | On meshlabs.cloud, if you allow analytics, Google Tag Manager and the Google Analytics tag it loads collect page views, approximate location derived from your IP address, browser and device information, referring page and interactions, using cookies or similar identifiers. If you do not allow it, they are not loaded. Whichever you choose, we keep an anonymous daily count of the choice (§07). |
04 Where the Data Comes From
We obtain personal data from the following sources:
- Directly from you, when you create an account, configure a node, submit a job, contact us or make a payment;
- Automatically, when your browser, desktop or mobile app, IDE extension, SDK client or node connects to the control plane, executes shards, exchanges signaling messages or completes jobs;
- From peer nodes, in the form of trustless validation reports, divergence measurements and reputation feedback;
- From third parties, including identity providers if you sign in using SSO, payment processors, identity-verification (KYC) providers, anti-fraud and anti-abuse vendors, public IP-reputation feeds and, when you enable Agentic Features, websites, APIs, MCP servers, skills and connected account providers with which the agent interacts.
- From your device and your instructions, when you grant access to files, folders, commands, applications or browser sessions, configure permissions or schedules, approve an Agent Action, or ask an agent to interact with local or connected resources.
05 Purposes & Legal Bases for Processing
| Purpose | Categories used | Legal basis (EU/UK GDPR) |
|---|---|---|
| Provide the Service: account management, scheduling, peer matching, shard distribution, signaling, job execution, result delivery. | Account, device, network, usage, job content and artifact access links. | Performance of a contract; consent (where required for sensitive client-side storage). |
| Operate the Compute Units (CU) and reputation systems, including oracle-based reward calibration and reputation scoring. | Contribution metrics, reputation, hardware attestation. | Performance of a contract; legitimate interests in operating a fair, sybil-resistant marketplace. |
| Multi-device access to Personal AI conversations: signaling the peer-to-peer connections between your signed-in devices, and, for conversations that are not Private, keeping their jobs in your History so they are available across your devices. | Personal AI conversations, account, network identifiers. | Performance of a contract. You choose per conversation, with Private, whether its jobs are kept in History. |
| Private jobs: keeping anonymous statistics (lengths, counts, sizes, timing) and non-content parameters to operate, price, size and reward the Service without storing content. | Usage and job metadata. | Performance of a contract; legitimate interests. |
| Asking you to approve an agent action on your phone, and telling you there that a task on your computer has ended or is waiting for you, when the mobile application is not open (push notifications). | Mobile push token, account. | Performance of a contract; consent to notifications given in your phone's settings. |
| Web search and page reading in the web application, where offered (searches are sent through our servers to our search provider, and pages are fetched from our network without your cookies). | Search queries, page addresses, account, network identifiers. | Performance of a contract. |
| Content moderation, where enabled: screening prompts and input images of a job before it runs, as some model licences require. | Job content, job metadata. | Legitimate interests in preventing abuse; compliance with the licence terms of the models you choose; legal obligation where applicable. |
| Diagnosing failures on your devices (node events, degradation reports, device benchmark history and, on test builds or where enabled, redacted desktop log extracts). | Device, network, logs and security data. | Legitimate interests in a reliable Service. |
| Personal space, where enabled: proposing and storing profile entries from summaries of your own non-Private jobs, and letting the model you use read them through tools. | Job content of non-Private jobs, personal space entries. | Consent (you can turn it off and delete entries). |
| Website analytics on meshlabs.cloud. | Website analytics, device, network identifiers. | Consent where required by law; otherwise legitimate interests. |
| Provide Agentic Features: interpret instructions, access user-approved resources, operate enabled tools and connected accounts, request and record permissions, perform Agent Actions, maintain action history, prevent abuse and support revocation. | Agentic content, permissions and actions; connected accounts and communications; account, device, network and security data. | Performance of a contract; consent where required for a particular permission, connected account, sensitive device access or storage; legitimate interests in security, abuse prevention and maintaining an auditable Service. |
| Network-aware scheduling: inferring which nodes are likely on the same local network (using the pseudonymized network-locality identifier and peer latency measurements) in order to place related work units on nodes with fast connections to each other. | Network identifiers, contribution metrics. | Performance of a contract; legitimate interests in efficient job placement. |
| Trustless validation and abuse prevention: redundant execution, canonical hashing, divergence statistics, sanction-screening, rate-limiting, fraud detection. | Job metadata, reputation, network identifiers, security data. | Legitimate interests; legal obligation (sanctions and AML). |
| Billing, invoicing, taxation, accounting, audit and bookkeeping. | Account, payment data, contribution metrics, CU transactions. | Performance of a contract; legal obligation; legitimate interests. |
| Service security, integrity and continuity: incident response, vulnerability management, intrusion detection, backups, disaster recovery. | All categories, on a need-to-know basis. | Legitimate interests; legal obligation. |
| Customer support and user communications. | Account, communications, job metadata. | Performance of a contract; legitimate interests. |
| Service improvement, analytics, capacity planning and benchmarking on the Oracle Node, primarily using de-identified or aggregated data. | Device, usage, job metadata, contribution metrics. | Legitimate interests; consent for non-essential analytics where required. |
| Compliance with legal obligations and protection against legal claims. | As required. | Legal obligation; legitimate interests; establishment, exercise or defense of legal claims. |
| Marketing communications, where you opt in. | Account, communications. | Consent (you can withdraw at any time). |
| Usage signals from agent tasks: finding and fixing failing tools, connectors and missing capabilities (§8.6). | Coded, content-free usage events; application, version and model. | Legitimate interests in a working Service; you can object at any time in Settings. |
| Improvement proposals you choose to send (§8.6). | The text you approved; your account only if you attach it. | Consent, given each time you press Send. |
06 Peer-to-Peer Data Flows (Important)
Mesh does not control Provider devices. The Service is designed with the architectural assumption that nodes are untrusted, and we apply trustless validation, quorum acceptance and reputation-aware scheduling to mitigate that risk. Despite these measures, network execution is not appropriate for regulated or confidential content unless an applicable agreement and Service configuration expressly permit it. See the Acceptable Use and Requester-Specific Terms in our Terms & Conditions. Available controls may include:
- anonymizing or pseudonymizing inputs before submission; and
- using a local-only or other approved execution mode rather than the peer network when working with sensitive content.
Presigned artifact URLs are bearer capabilities rather than session-bound links. Possession of the complete URL may permit access without a further Mesh authentication check until the signature expires. Do not publish or forward such URLs, and treat them as confidential while valid.
6.1 Where a Request Runs, and What "Private" Means
Each request runs in one of these places, and the application shows which one you have selected:
- This device. A model in the desktop application's local runtime (including its offline mode) runs on the computer you are using. Its content does not reach our servers or other devices. A request in the web application runs as a job, even when the device that runs it is your own browser tab.
- Your own devices. A job the control plane may run only on devices registered to your account. The agent uses this for content from sensitive sources (see §08).
- Your clusters. Your personal cluster, and any cluster you own or may run work on. Devices that other people connected to a cluster you own, for example through its invitation link, can receive its jobs.
- The Global Network. Devices of unrelated Providers.
In the last three cases the job passes through our control plane, and its inputs and outputs may be held in our temporary stores and object storage while it runs. Private decides what we keep afterwards (§03, Job content); it does not change where the job runs or which devices see its content while it runs.
08 Agentic Features Data Flows
Agent processing may occur locally on your device, but enabling a website, connected account, API, cloud model, MCP server, skill or other network tool may transmit selected prompts, files, page content, command output, tool inputs, communications or other data to that third party. Information intentionally published, uploaded or sent by an Agent Action becomes available to the selected public audience, platform, account or recipient and may be copied or retained by them. Mesh cannot delete or retrieve copies held by those recipients.
The application may store agent transcripts, action and approval history, granted folders, capability settings, installed MCP servers and skills, schedules and a dedicated browser profile locally under your signed-in account profile. AI model files may be shared across accounts on the same device. Application-level separation does not prevent a person with operating-system or administrator access from accessing unencrypted local files or browser-profile data.
A dedicated in-app browser may retain cookies and authenticated sessions so connected sites remain signed in. Sensitive credentials entered during a human-takeover flow are intended to remain between you, the browser and the relevant site rather than being provided to the AI model. Access tokens and browser sessions are nevertheless sensitive data. You should grant the narrowest available permissions, review requested actions, disconnect integrations you no longer use and use available controls to stop Agent Actions and revoke grants.
Permission and action records are processed to enforce your choices, display an action history, investigate failures and abuse, and provide security and support. If an action is approved from a paired device, we may also process the approving device, time and decision as part of that record.
8.1 Your Own Browser (MESH Agent Browser Extension)
If you install the MESH Agent Browser extension and pair it with the desktop application, the agent can act in your own Chrome browser, where you are already signed in to websites. The extension works only on a tab it opened itself, only on a site you allowed in the desktop application, and only when the paired desktop application sends a command. On that tab it can read the page's text and controls, click, and type into fields that are not passwords, payment-card fields or one-time codes. It does not read cookies, saved passwords, browsing history or other extensions, does not take screenshots, and does not run code downloaded from the network.
It does not read your other tabs, with one exception: when you click a link to one of its tabs in the MESH desktop app and that tab has since closed or moved to another page, it asks Chrome for a tab open at the link's address in the same Chrome profile and brings that tab to the front. It reads nothing on that tab, tells the app only whether it found one, and sends nothing else about your tabs anywhere. Chrome gives the extension only the tabs whose address begins with the link's address, and the extension compares each one with the link before it shows it.
The extension sends what it reads only to the MESH desktop application on the same computer, through Chrome's native messaging channel; the extension itself sends nothing to our servers or to anyone else. The desktop application then uses the page text as part of the agent's conversation, and it goes wherever that conversation's model runs (§06.1). Text from sensitive sites stays on your own devices (§08.2). The extension uses what it reads only to carry out the task you gave the agent; it is not used for advertising, sold, or used to train AI models. Our use of information the extension handles adheres to the Chrome Web Store User Data Policy, including the Limited Use requirements.
8.2 Sensitive Sites and Sources
The desktop application keeps a list of sensitive sites (for example messaging, email, social networks, banking, payments, health, government and identity, and cloud documents). On those sites, in the in-app browser or in your own browser, every outgoing action (such as sending, posting, paying, deleting or sharing) asks you each time, and the page content may be read only by a model that runs on this computer or on your own devices, never on other people's devices. The same rule applies to mail from the Gmail connector. Requests that carry such content are sent as Private jobs. The activity record keeps the action, the site and the size of any text, not the text itself.
WhatsApp Web. If you ask the agent to work with WhatsApp, it opens WhatsApp Web in your own browser (through the extension, §8.1) or in the in-app browser, signed in with your own WhatsApp account; we have no WhatsApp account or connection of our own. It reads the text of the WhatsApp Web pages it opens for your task, such as your chat list and the chats you ask about, including the names or phone numbers and messages of the people you talk with, and it can type a draft into the message box; it never sends, and you press Send. As on every sensitive site, what it reads is processed only by a model on this computer or as a Private job that our control plane runs only on devices your MESH account owns, so we do not store it in our database: the job's inputs are held in temporary memory for at most 2 hours and its outputs for up to 15 minutes (§11). Like any agent task, it becomes part of the conversation in your chat history on your desktop and the devices that sync it, until you delete it. It is not saved to long-term memory or offered in a feedback proposal (§8.6), and the activity record keeps the action and the site, not the text.
8.3 Gmail and Other Google User Data
If you connect Gmail to the desktop application, you sign in with Google using an OAuth client from a Google Cloud project you configure in the application's settings. The application then uses the Gmail API, from your computer, to search and read messages and threads, create drafts, send messages you approve, and, if you allow it, label and archive messages. The OAuth tokens are sealed with your operating system's keychain on your computer and are never sent to our servers. Mail content is used only to answer the request you made, may be read only by a model that runs on this computer or on your own devices, is sent as a Private job when it runs on a cluster, and every send or filing asks you first. Disconnecting revokes the grant with Google and deletes the stored tokens.
MESH's use and transfer of information received from Google APIs to any other app will adhere to the Google API Services User Data Policy, including the Limited Use requirements. The use of information received from Google Workspace scopes will also adhere to the Google Workspace User Data and Developer Policy. In particular, we do not use Google user data for advertising, do not sell it or transfer it to data brokers, do not use it to train or improve AI or machine-learning models, and do not allow people to read it unless you ask us to (for example in a support request), it is necessary for security, or the law requires it.
8.4 Other Connectors (MCP Servers)
You can connect the desktop agent to Model Context Protocol (MCP) servers, such as the ones Linear, Notion, Atlassian (Jira and Confluence) and GitHub run, or servers you add yourself. The server receives the requests and arguments the agent sends to it and returns results under its own terms and privacy notice. Sign-in tokens are sealed on your computer and revoked when you disconnect. Each server has a setting for where its results may go: this computer only, your own devices, your clusters (including devices other people connected to them), or anywhere including the Global Network. Linear, Jira, Confluence and GitHub default to your clusters; Notion and servers you add yourself default to your own devices; only you can widen a setting.
8.5 Scheduled Tasks and Web Tools
If you approve a schedule, the desktop application sends its prompt as a new turn at each run time while the application is open, without asking again. A scheduled run cannot write, edit or delete files or run commands; any other action you have not already allowed asks first, and is refused if nobody answers within five minutes. The desktop agent's web search and page reading run from your computer, using the search service whose key you entered (for example Brave or Google) or, without a key, public sources such as Google News, DuckDuckGo and Wikipedia; those services receive your search terms and your IP address.
8.6 Feedback to MESH: Usage Signals and Improvement Proposals
The agent can tell us what does not work, in two ways. Neither ever sends anything in your name without you, and neither is used to train AI models.
Usage signals (on unless you turn them off). After an agent task, the desktop application records what broke or was missing, as codes from a fixed list: which built-in tool or which connector (by vendor, for example "Figma", or "custom") failed and the kind of error (for example "timed out" or "sign-in failed"); a capability the model asked for that MESH does not offer (a category such as "spreadsheet", never the name the model wrote); how the task ended; coarse duration and step-count ranges; the application, its version and channel, the operating system, the catalog model and whether it ran on this computer, your devices, your cluster or the network; and how many times each happened. It never includes what you or the model wrote, your files, file paths, web addresses, tool arguments or results, error messages, or any account, device or install identifier, and our servers reject a signal that contains anything outside that fixed list. Signals are sent at most once an hour, with no sign-in credential, and dropped when you are offline. We store only daily totals per combination of those codes, so a stored signal cannot be traced back to you. Our legal basis is our legitimate interest in finding and fixing failures in the Service; you can object at any time with Settings → Usage signals on the desktop, the website or the phone app. Turning it off while signed in is recorded on your account, and every desktop signed in to it stops sending; signed out, it applies to that desktop.
Improvement proposals (only when you press Send). When a task hits a real gap (a tool or connector failing again and again, a capability MESH lacks, or you saying it should work better), the desktop application may show you a short draft proposal. It is written from the signals above and, when you asked for something to work better, the sentence you wrote; before you see it we remove email addresses, file paths, the paths and queries of web addresses, phone numbers, IP addresses, handles and the names we can detect, and what you see is exactly what would be sent. You can edit it, choose "Send to MESH" or "Don't send", and choose to send it anonymously (the default) or to attach your account so we can follow up. Nothing is sent if you do not answer. A draft is never offered for a Private conversation, for content limited to this computer, your own devices or your clusters, for Gmail or other connector results, or for content from sensitive sites (§8.2–8.4), and our servers refuse a proposal the application marks as coming from one; you can still write to us in your own words with Settings → Send feedback. An anonymous proposal is stored with no account and no IP address; an attached one is stored with your account. Proposals are labelled "sent by the user via the MESH agent" (or "via the feedback form"), are read by our staff to decide what to build, and, if we track one in our issue tracker, we describe it in our own words without your text or account. Our legal basis is your consent, given each time you press Send; an attached proposal can be deleted on request (§13) and is deleted with your account.
Like any request to our servers, these pass through our load balancer and web firewall, whose logs may record your IP address for the periods in §11; to limit how often one network can send, our rate limiter keeps the IP address for about an hour.
8.7 Your Control
You are in charge, and the agent acts for you: only on a task you give it or a schedule you approved. Before it first opens a site in a browser, changes your files, runs a command that can change something, uses a tool server or starts a generation job on the MESH network, it asks you on a card, and you choose whether your answer covers once, this session or always. On a site you allowed, it then reads and moves between pages without asking again, and a command you allowed in a folder can change or delete files in it. Some actions ask you on their own card every time, whatever you allowed before: anything on any website that could send, post, pay, transfer, delete, share, file or change a setting, or submit a form other than a search; sending or filing email through Gmail; deleting a file with the agent's file tools; a git push or a rewrite of git history; and installing packages, a runtime, a skill or a tool server. A scheduled task does none of those on a website and says what it would have done. On WhatsApp Web the agent never sends: it writes a draft, and you press Send. On sites whose terms forbid agents, such as LinkedIn, Discord and amazon.com, it does nothing at all. You can press Stop to end a task at any time (it does not undo what is already done), revoke remembered approvals in the app, and disconnect the browser extension or Gmail.
10 International Data Transfers
Mesh is a Delaware corporation with its principal mailing address in the United States. Our control plane, database and object storage are hosted by AWS in the United States (the us-east-2 region), and backups may be kept in another AWS region. Our other service providers and peer nodes may be located throughout the world, including in destinations that may not provide the same level of data protection as your home jurisdiction. Peer-to-peer execution may transfer job content directly to the countries in which assigned Provider nodes are located.
Where the EU/UK GDPR or Swiss data-protection law requires a transfer mechanism, we rely as applicable on the European Commission's Standard Contractual Clauses, the UK International Data Transfer Addendum or Agreement, the Swiss adaptations to the Standard Contractual Clauses, adequacy decisions, or another lawful mechanism. We apply supplementary measures where appropriate, such as encryption in transit and at rest, access controls, minimization and transfer-risk assessments. You may request information about applicable safeguards by contacting privacy@meshlabs.cloud.
11 Data Retention
We keep personal data only as long as necessary for the purposes described in this Policy and to satisfy our legal, accounting and reporting obligations. Indicative retention windows (which may be adjusted to reflect operational and legal requirements):
- Account data: for the life of the account. An account you delete yourself is deleted 30 days after your request, as described in §11.1; an account closed in another way, up to 24 months after closure.
- Uploaded job inputs and transient artifacts: files you upload for a job, and the intermediate files devices upload while working on it, are deleted automatically one day after they are stored. Object storage applies that rule once a day, so such a file is permanently deleted within 3 days.
- Private jobs: inputs held in temporary memory only while the job runs (at most 2 hours) and outputs for up to 15 minutes; neither is written to our database. Files the job uploads or produces follow the previous item and are permanently deleted within 3 days. Their anonymous statistics follow the job-metadata period below.
- Final generated outputs and saved jobs: output records and downloadable artifacts, and the jobs of your saved (not Private) conversations and generations, are retained for 365 days to provide job History and result delivery, and then permanently deleted, unless you delete them earlier. When you delete a job from History, its saved output files are deleted at once and any remaining intermediate files within 3 days. Artifact files may become unavailable before the associated job record, so you should download outputs you wish to retain. Jobs submitted through the developer API relay are kept for 24 hours. The app shows these periods beside the Private choice.
- Deletion is permanent: when a file is deleted from our object storage, no earlier version of it is kept, it is not copied to a backup, and our content delivery network neither serves nor caches job files.
- Data on your devices: Personal AI conversations, generated media kept by the desktop application (up to 20 GiB by default), memory, schedules and connector tokens stay on your devices until you delete them, delete the conversation (which also removes it from your other desktops), disconnect the connector, remove the account's data from the application, or uninstall the application.
- Personal space (where enabled): until you delete entries or use "Delete everything", or your account is deleted.
- Presigned artifact URLs: input download links are generally issued for up to 24 hours and output links are generally issued for shorter periods and may be refreshed while the underlying artifact remains available.
- Job and contribution metadata, CU ledger entries: up to 7 years for accounting, audit and tax compliance.
- Agentic Features data stored locally: agent transcripts, action and approval history, granted roots, capability settings, schedules, installed MCP servers and skills, and the per-account browser profile are retained on the device until you remove the applicable data, integration, account profile or application, subject to available product controls and operating- system behavior.
- Agentic data held by third parties: information sent to websites, connected accounts, MCP servers, skills, message recipients or public audiences is retained under those parties' policies and cannot necessarily be deleted through Mesh.
- Security and abuse logs: typically 90–365 days (audit records 365 days; application logs typically 30–90 days).
- Operational telemetry: node metrics 7 days, node runtime events 30 days, diagnostic snapshots 24 hours, uploaded desktop log extracts 7 days, and sign-in sessions up to 7 days after their last refresh (records of revoked sessions 30 days).
- Feedback to MESH (§8.6): usage signals are kept only as daily totals, for 400 days; improvement proposals for 365 days, and a proposal with your account attached is deleted with your account.
- Backups: rolled off on a defined backup-retention schedule (typically up to 35 days).
Where you exercise your right to erasure (§13), we will delete or de-identify personal data unless we are legally required or permitted to retain it.
11.1 Deleting Your Account
You can delete your account yourself under Settings → Delete account on the website and in the MESH phone app, or by emailing privacy@meshlabs.cloud. Before we act we ask you to prove again that it is you: your password, or, for an account you sign in to only with Google, a new Google sign-in completed within the last few minutes. Being signed in, or knowing the account's email address, is not enough.
- At once: the account is deactivated, every session on every device is signed out, your personal API key and public LLM endpoint stop working, computers can no longer join the network under your account, the phone notification tokens registered to your account are removed, and we email your address with the deletion date and how to cancel.
- After 30 days: we delete your saved jobs and their outputs (job History), your personal space, the devices, sign-in identities, password, sessions and notifications registered to your account, your desktop log uploads, survey answers, the clusters you own and your memberships in other clusters, and we replace your email address and name on the account record with an anonymous identifier. Until then you can cancel the deletion by contacting support.
- What we keep: the payment and legal records linked to the account, under the anonymous identifier rather than your email address: the credit ledger, payments, refunds, withdrawals and their payout details, compute contributions and earnings, the Stripe customer and connected-account references (Stripe keeps its own records under its policy), the versions of these Terms, this Policy and the in-app disclosures you accepted, and acceptances of model-specific licence terms (which also keep the email address shown when you accepted). Security audit records keep your email address for up to 365 days. Backups roll off as described above.
- Money on record: if the account has a balance or a withdrawal in progress, we ask you to confirm that you know it stays on record. Deleting the account does not cancel or forfeit it; ask support to be paid, within the period in Terms §18.
- Organizations: an organization's jobs, keys and billing belong to the organization and are not deleted with a member's account. The last admin of an organization that has other members must make another member an admin first.
- Data on your devices (Personal AI conversations, memory, connector tokens) is not on our servers; remove it in the application or by uninstalling it.
12 Security Measures
We implement technical and organizational measures appropriate to the risks presented by our processing, including:
- encryption of data in transit (TLS) and at rest where applicable;
- role-based access control with the principle of least privilege, enforced at the API, service and database layers and segmented by organization (RBAC);
- cryptographic integrity verification of model assets (e.g., SHA-256 content addressing);
- trustless validation, redundancy and quorum acceptance for compute results;
- connector, browser-pairing and sign-in tokens sealed on your devices with the operating system's keychain or equivalent, and passwords, payment-card fields and one-time codes withheld from the AI model;
- logging, monitoring, intrusion detection and anomaly detection;
- for Agentic Features, permission scopes and revocation controls, folder boundaries, action and approval history, per-account browser profiles, and human-takeover flows for sensitive interactions, where applicable;
- secure-development practices, code review and dependency-vulnerability management;
- a documented incident-response process, including notifications to regulators and affected users where required.
No security measure is perfect. If you suspect a vulnerability or breach, please contact security@meshlabs.cloud promptly.
13 Your Privacy Rights
Subject to applicable law (including the EU/UK GDPR, Swiss and U.S. state privacy laws, and the CCPA/CPRA), you may have the following rights:
- Access. Confirmation of whether we process personal data about you and a copy of that data. See Article 15 GDPR.
- Rectification. Correction of inaccurate or incomplete data. See Article 16 GDPR.
- Erasure ("right to be forgotten"). Deletion of personal data, subject to exceptions (e.g., legal obligations, defense of claims). See Article 17 GDPR.
- Restriction. Restriction of processing in defined circumstances. See Article 18 GDPR.
- Portability. Receipt of personal data in a structured, commonly used, machine-readable format and transmission to another controller, where technically feasible. See Article 20 GDPR.
- Objection. Objection to processing based on legitimate interests, including profiling. See Article 21 GDPR.
- Withdrawal of consent. Where processing is based on consent, you may withdraw it at any time without affecting the lawfulness of prior processing.
- Non-discrimination. We will not discriminate against you for exercising any of these rights.
- Lodging a complaint. You may lodge a complaint with your local data-protection authority.
To exercise your rights, including a request to delete your account and its data or to receive a copy of it, email privacy@meshlabs.cloud from the address associated with your account, or use the in-product privacy controls where available (for example Settings → Delete account, §11.1; the Private setting, deleting conversations, History items and memory records, clearing the agent browser's data, disconnecting connectors, and exporting or deleting your personal space). We may need to verify your identity before responding. We will respond within the time periods required by applicable law (typically 30 days under GDPR, extendable by 60 days for complex requests; 45 days under CCPA, extendable once).
14 Children
The Service is not directed to, and we do not knowingly collect personal data from, individuals under the age of 18 (or the age of legal majority in their jurisdiction, whichever is greater). You may not enable Agentic Features if you are under eighteen (18). If you believe a child has provided us with personal data, please contact privacy@meshlabs.cloud and we will take appropriate steps to delete it.
15 Automated Decision-Making & Profiling
The Service performs a number of automated processing operations that may affect you, including: scheduling jobs to peer nodes (using device capability, network locality, latency and reputation inputs), accepting or rejecting compute results based on quorum and divergence statistics, computing reputation and CU, and detecting fraud, abuse and Sybil behavior. If you enable Agentic Features, AI models also interpret your instructions, select tools and propose or perform actions within the permissions you configure. Those actions may affect files, accounts, communications, publications, purchases and third parties.
We endeavor not to engage in automated decisions that produce legal effects or similarly significantly affect you within the meaning of Article 22 GDPR. Where such decisions occur (for example, an account suspension based primarily on automated abuse signals), you have the right to obtain human review, express your point of view and contest the decision by contacting privacy@meshlabs.cloud.
This statement about decisions Mesh makes concerning you does not mean that an Agent Action you instruct or permit cannot have legal or similarly significant consequences. You should use human review and confirmation for consequential actions involving contracts, purchases, employment, finance, healthcare, government services or other people's rights.
16 Use of Data for AI Training & Improvement
We do not use the content of Requester prompts, source media, source code, Personal AI conversations (whether stored locally, relayed, or stored server-side at your request), agent transcripts, selected file or page content, command or MCP output, communications, Agent Actions or generated outputs to train Mesh's own foundation models without a separate, opt-in agreement with you. We may use de-identified or aggregated job metadata (for example, model identifiers, latency distributions, hardware-class distributions, average shard memory consumption) to evaluate, benchmark and improve the Service, including to recalibrate the deterministic complexity profile used by the Oracle Node, optimize job scheduling and node matching, plan network capacity and refine validation mechanisms. We do not describe data as anonymous unless it has been processed so that it no longer relates to an identified or identifiable person under applicable law.
We never use Google user data (§08.3) or content read through the browser extension (§08.1) to train or improve AI or machine-learning models. Outputs of models whose licence forbids using them for training are marked as such when they are stored.
17 Region-Specific Disclosures
17.1 European Economic Area, United Kingdom and Switzerland
The legal bases on which we rely are listed in §05. You may contact our team at privacy@meshlabs.cloud for any data-protection enquiry. Where Article 27 of the EU or UK GDPR requires us to appoint a local representative, the representative's identity and contact details will be published in this Policy.
17.2 Israel
Our processing is governed by the Israeli Privacy Protection Law, 5741-1981 and its regulations, including the Privacy Protection (Data Security) Regulations, 5777-2017. Subject to applicable law, you may have rights to access information about you in a database and to request correction or deletion of information that is inaccurate, incomplete, unclear or outdated, including under Sections 13 and 14 of the Israeli Privacy Protection Law. Contact privacy@meshlabs.cloud to exercise those rights.
17.3 United States, including California
Residents of U.S. states with applicable comprehensive privacy laws may have rights to access, correct, delete or obtain a copy of personal data; opt out of certain sales, targeted advertising or profiling; and appeal a denied request. To the extent the CCPA/CPRA applies, California residents may also request to know the categories and specific pieces of personal information we collect and limit certain uses of sensitive personal information. We do not sell personal information or share it for cross-context behavioral advertising. During the preceding 12 months, we may have disclosed the categories described in §03 for the business purposes and to the recipient categories described in §09; the categories collected, sources, purposes and recipients are described in §§03–09. We have not sold personal information or shared it for cross-context behavioral advertising. Some information you ask the Service to handle is sensitive personal information under the CCPA (for example account credentials, financial account details for payouts, or the contents of email and messages the agent reads for you); we use and disclose it only to provide the services you request and for the other purposes the CCPA permits, and not to infer characteristics about you. To exercise rights or submit an appeal, contact privacy@meshlabs.cloud. You may designate an authorized agent, subject to legally permitted verification, and we will not discriminate against you for exercising applicable rights.
17.4 Other Jurisdictions
We will honor applicable rights granted under other privacy laws (e.g., Brazilian LGPD, Canadian PIPEDA, Singapore PDPA). Where local law requires specific disclosures or additional rights, those will apply in addition to the rights listed in §13.
18 Changes to This Policy
We may update this Policy from time to time. When we make material changes, we will provide reasonable notice (e.g., by email or in-product banner) and update the "Last Updated" date above. Your continued use of the Service after the effective date of the updated Policy constitutes acceptance of the changes.
Version history. Version 1.7, effective November 15, 2026, adds a second use of phone notifications to §3 and §5: telling you that a task on your computer has ended or is waiting for you. No other section changed.
19 Contact & Complaints
For questions, requests or complaints regarding this Policy or our data practices:
Mesh Labs Cloud, Inc.
Attn: Privacy
2810 North Church Street
Wilmington, Delaware 19802, United States
Privacy: privacy@meshlabs.cloud
Security: security@meshlabs.cloud
Legal: legal@meshlabs.cloud
If you have an unresolved concern, you may also contact your local data-protection authority or other competent regulator.