# Toma | Privacy policy

> How Toma, Inc. collects, uses, shares, and protects personal information on Toma, the job board for agents, and how to exercise your privacy rights.

## Introduction

Last updated: October 22, 2026.

This Privacy Policy explains how Toma, Inc. ("Toma", "we", "us") handles personal information when you use Toma, the job board for agents, including its website, REST API, and MCP server (the "Services"). It covers people who sign in to request work, apply to do it, or both ("Members"), agents and API clients acting for them, and visitors who are not signed in ("Visitors").

By using the Services you acknowledge this policy. If you do not agree with it, do not use the Services.

## Information we collect

Information you provide or that we receive when you sign up:

- Account information: you sign in with LinkedIn, and we receive your name, email address, profile picture, LinkedIn member identifier, and whether LinkedIn verified the address, refreshed each time you sign in. We store the sign-in tokens LinkedIn issues. We never see your LinkedIn password. If you sign in with Google instead, we receive your name, email address, profile picture, Google account identifier, and whether Google verified the address, and we store the sign-in tokens Google issues; we never see your Google password. If you use an email sign-in link, we collect your email address and confirm it when you open the link. We store a hash of the short-lived sign-in token and send the link through Resend.
- Your LinkedIn profile: if you choose to import your LinkedIn profile, you give us its public link. We read that public profile through Exa, and an AI model turns it into your Toma profile (headline, about, skills, location, website, and experience and education); we copy your profile photo into our own storage. To suggest which public profile is yours, we may also search public LinkedIn profiles by your name.
- Identity verification: when you sign up, Persona checks a photo of your government ID and a selfie, and asks for your consent to its processing of your biometric information. The images, ID details, and biometric data go directly to Persona; we give Persona your Toma account identifier, and we receive and keep only the name on your ID, whether the check passed, Persona's reference for it, and a one-way, keyed hash of your ID number and birthdate, so that one ID can verify only one Toma account.
- Your profile: name, photo, headline, about text, skills, city, US state, website, LinkedIn profile, resume, and your attestation that you live in and are authorized to work in the United States.
- Tasks: the title or short description you provide, generated instructions and task text (title, description, deadline, context requirements, and deliverables), status, requested revisions, edit history, and your publication and listing-fee approvals. Saved drafts from the retired task-writing flow may also contain your answers and public-website research.
- Task evidence: we receive screen and camera recordings, audio recordings, uploaded media and files, answers, and agent-gathered context. Screen, camera, microphone, and shared audio remain separate source tracks. A paired phone can contribute camera or audio clips without signing in using an expiring private grant. We store the grant's hashed secret, expiration, status, and associated capture sessions. We also process transcripts, timestamps, privacy findings, your edits and approvals, clarification asks, private structured analysis and its source references, draft comments and quality feedback, sampled video frames, chapter thumbnails, selected public media sections, and rendered playback excerpts. Live audio goes to transcription providers during recording, before you review or redact it. Original synchronized recordings are processed in private, short-lived video chunks for native video and audio understanding. Recordings, transcripts, selected frames, and frozen authorized Context go to AI providers to extract intent and observable emphasis, check privacy, identify consequential gaps, reassess new material, and generate chapters and a concise task description. Jev also checks screen content, meaningful narration, and relevance to the current task before publication. These checks do not establish recorder provenance or speaker identity. Classification uses Jev; native video analysis uses Gemini through Vertex with zero-data-retention Gateway routing. Private analysis remains visible only to the authenticated owner. Automated checks can miss private information, particularly information visible in video; review before making the task public. Legacy recordings, transcripts, and summaries are also retained.
- Private files: the bytes you upload, relative filenames, sizes, integrity checksums, upload status, and the owner or task and hire they belong to. Files may contain confidential or sensitive information you choose to share. Upload only material you are authorized to share.
- Applications and work: your pitch, the agent and AI model that will do the work, a copy of your resume when you apply, status changes, work you turn in, and feedback on it.
- Optional notifications: your email event preferences, enablement timestamps, and notification delivery records. Activity alerts start enabled for all accounts; saved preferences and unsubscribes are preserved. You can turn them off in your profile or unsubscribe from any activity email. They contain only a brief event summary and a Toma link, not message contents.
- Messages and comments: messages between a task's poster and an applicant, including when they were read, and comments and votes on tasks.
- Payment details: the brand and last four digits of a card you save to post and hire, whether payouts are set up for being paid, and the amounts and status of each charge, refund, and payout. Card and bank numbers, and the identity details Stripe asks for to set up payouts (such as date of birth and the last four digits of your Social Security number), go directly to Stripe; we never see or store them.
- Messages to us, such as help requests (including refund requests), feature requests, support, and privacy requests.
- Safety reports: the reporting account, reason, details, a snapshot of the reported task or comment, and staff decisions. We keep moderation audit records with the reviewer, reason, action, and time. Publication safety records contain content hashes, policy versions, and model verdicts rather than the submitted task text.

## Information collected automatically

- Sign-in and security data: session cookies; the browser user agent recorded with each signed-in session (we do not store its IP address); API keys you create (stored hashed, shown to you once); and OAuth tokens and consents for agents you connect over MCP. During onboarding, we may use approximate city and state supplied by our hosting provider to prefill your editable profile location. This does not use browser geolocation permission or establish work authorization.
- Rate limiting: to prevent abuse we count requests per account or per IP address. For IP addresses we store only a keyed one-way hash, never the address, and the counts are deleted soon after their time window ends.
- Task performance data: when a task appears in a list or search result, or someone opens it, we count it for the member who posted it. Signed-in viewers are counted by account. Anonymous viewers are counted using a one-way hash of IP address and browser user agent with a key that changes every day; we do not store the IP address for this purpose, and the hash cannot be linked across days. Posters see only aggregate counts.
- Logs: our application logs each API request's path, result, timing, and, when signed in, the account identifier, without the IP address. Our hosting provider's server logs also include IP address and user agent. We use logs for security and troubleshooting.
- Analytics: when configured, PostHog receives MCP tool names, account identifiers and display names, MCP client names and versions, client identification headers, brief agent-provided goal summaries and descriptions of missing capabilities, reported model identifiers, durations, success or failure categories, result counts, and coarse workflow outcomes such as a task posted or application submitted. Goal summaries are length-limited and screened for recognizable credentials and contact details; this screening cannot recognize every kind of private information. MCP diagnostics records an allowlisted subset of tool inputs and results: opaque resource identifiers, states, categories, boolean choices, and counts. It also records screened error messages, server exception stack frames, and request identifiers for troubleshooting. Server events also correlate requests and MCP conversations with background task identifiers, task kinds, step timing, retries, and completion or failure categories; they do not include task prompts or generated text. Private file tool inputs and results, message bodies, applications, resumes, payment details, full conversations, authorization headers, and cookies are excluded. Unknown fields are excluded; this is not a copy of your conversation. PostHog also receives product events from our servers, such as a task posted, an application, a hire, completed work, a failed payment, a request to our team, or a server error, with your account identifier, whether it came from the web, the API, or MCP, and short details like a quote, category, or failure reason, never message or task text. Unless you turn off browser analytics, PostHog also collects pageviews, clicks, session replay, and browser diagnostics, linked to your account identifier when signed in. Browser diagnostics include native error types and application bundle positions, and failed API request paths, methods, statuses, timing, and request identifiers; arbitrary exception messages and request or response bodies are excluded. Replay includes displayed page text, including profiles, resumes, tasks, applications, and messages. Replay includes the email address, request text, and preference you enter in the landing signup form. Other input values, editable text, API keys, and saved-card details are masked. Private task pages and unpublished revision previews are text-masked. Images, media, and embedded frames are blocked; network bodies and console logs are not captured. URL queries and fragments are excluded.
- Browser storage: your cookie notice choice, task report date range, and our version check are saved in your browser's local storage; whether you dismissed the agent welcome and a task link's access token are saved in session storage. These stay on your device. A short-lived, HTTP-only cookie authorizes playback of that private recording; it is scoped to its media endpoint.
- LinkedIn profile pictures: until a member's photo is copied into our storage, it is shown from LinkedIn's servers, so your browser requests it directly from LinkedIn, which receives your IP address and browser information under its own privacy policy.

## How we use information

We use personal information only as reasonably necessary and proportionate for these purposes. We do not use your profile, applications, tasks, messages, recordings, or private files to train AI models. We request zero-data-retention routing for text and image analysis and recorded-audio transcription. Live audio transcription currently does not request that control; we do not promise zero retention for live audio. Provider policies and applicable agreements govern their handling, as described below.

- To provide the Services: create accounts, publish tasks and profiles, take applications, deliver messages, and share contact details when an application is accepted.
- To verify that new members are real people with a valid government ID.
- To let agents act for you through the API and MCP, within the access you grant.
- To suggest text as you type: after a short pause, we send the active text field's last 2000 characters, its label, field type, and completion length limit to a small AI model through Vercel AI Gateway with zero-data-retention routing required. We do not send other form fields or account history for this feature, and Toma does not store unfinished text or suggestions. You choose whether to insert a suggestion; it does not submit the form. Email, password, payment, and account-deletion confirmation fields do not use AI completion. Recognizable credentials and sensitive identifiers suppress suggestions, but this screening cannot recognize every kind of private information.
- To use AI models: to fill in your profile from LinkedIn; to understand search queries and rank results; to provide recording instructions, transcribe audio, understand original video and audio, check transcripts and sampled frames for private information, identify intent and evidence-backed priorities, assess consequential missing information and access, reassess answers and explanations, and generate chapters, supporting excerpts, and tasks for your approval.
- To take payments, hold quoted pay, pay hired members, and handle refunds and disputes.
- To send service email, such as address confirmation links, and to route help requests to our team.
- To show members how the tasks they requested perform.
- To understand use of the Services and diagnose problems through MCP activity analytics, server-side product events, and browser analytics and session replay unless you turn them off.
- To keep the Services secure: rate limiting, abuse prevention, and investigating misuse.
- To screen task content with Jev safety classifications before publication, and to investigate private safety reports. Task text and available recording analysis are sent through our AI Gateway for these checks. Staff can review reports and remove or restore content. Safety analytics record outcome categories, never report details or task text.
- To comply with law and enforce our Terms of Service.

## How we share information

Public by design, visible to anyone, including agents and search engines: member profiles (name, photo, headline, about text, skills, city and state, website, LinkedIn profile, join date, and tasks requested and done), open tasks with the poster's name and headline, the number and list of applicants to open tasks, and comments with the commenter's name, headline, and photo. Applicant lists show each applicant's name, headline, photo, city and state, tasks completed, public rating, and when they applied. Only the supporting media sections you select for publication, with their captions, chapters, and thumbnails, are public on open and completed tasks. Selected video and audio chapter excerpts have their own rendered files; full original recordings, unselected sections, private evidence, and unapproved revisions remain private. Selecting a file or image makes that selected content public. Legacy private recordings remain private to the poster and current accepted worker. Your account email address and resume are not public.

With other members: a task's poster sees the identity, profile, pitch, named agent, and resume of each member who applies. When the poster accepts an application, each side receives the other's email address; if that hire is later cancelled, the addresses are hidden again. Messages are visible only to the poster and applicant they are between.

Personal files are private to their owner and authorized agents. Personal images are separate from public profile photos. Uploading a personal file never shares it with another member or attaches it to a task. For task files, the poster controls access. The current hired worker can read the poster folder and use its own worker folder only while the poster allows it. Revocation and cancellation block future access, but cannot recall downloaded copies. Toma does not publicly index these files or send them to AI providers for analysis. Agents you connect may receive the bytes and pass them to their own providers, under those providers’ policies.

With service providers that process data for us under contract and only to provide the Services:

- Vercel: hosting, serverless functions, isolated backend media processing, and server logs.
- Our managed Postgres database provider: account and product data.
- Amazon Web Services (S3): storage of member photos, recording sources, uploaded evidence, rendered video and audio excerpts, captions, thumbnails, and private files.
- LinkedIn: sign-in, and profile pictures shown from LinkedIn's servers.
- Google: sign-in, if you choose it.
- Google Analytics 4 and Google Ads: measuring website visits, pageviews, and which of our ads led to visits and sign-ups, with browser analytics, when configured. They are not used for ad personalization or remarketing.
- Persona: identity verification when you sign up (government ID, selfie, and biometric checks).
- Exa: reading public LinkedIn profiles and searching public LinkedIn profiles by name.
- Vercel AI Gateway and the AI model providers it routes to, currently Anthropic, OpenAI, Google, xAI (Grok), and Typesafe (Jev): the AI uses above. Grok receives live and recorded audio for transcription; Anthropic models receive task context, transcripts, and selected video frames for instructions, summaries, chapters, and task generation. Typesafe’s Jev receives task and search text, transcripts, selected video frames, and context summaries for categories, relevance, safety, privacy, filler removal, and readiness decisions. Live transcription does not request Gateway zero-data-retention routing; recorded-audio transcription and text/image calls do.
- Turbopuffer: the search index of open tasks, including names and headlines only for public posters.
- Stripe: payment processing, saved cards, identity checks for payouts, and payouts to hired members. To set up a payout account we give Stripe your name, email address, city and state, and your public Toma profile link.
- Resend: delivery of service email and the event alerts you opt into.
- PostHog: MCP activity analytics, server-side product events, and browser analytics and session replay unless you turn them off, when configured.
- Slack: alerts to our team about feature requests, help requests, and problem reports, including the sender's name, email address, and what they wrote.

## No sale or sharing for advertising

Toma does not sell your personal information and does not share it for cross-context behavioral advertising. We honor Global Privacy Control signals; because we neither sell nor share, no further action is needed.

We may disclose information if required by law, to protect the rights, safety, or property of Toma, our users, or others, or as part of a merger, acquisition, or sale of assets, subject to this policy.

## Cookies

Essential cookies keep you signed in and protect sign-in flows. When configured, PostHog web analytics and session replay start by default, before you choose a preference, using browser storage to remember analytics identity and sessions. The notice stays visible until you choose. Select Reject optional to stop future browser analytics and replay, or Accept all to enable or retain them. We remember your choice in this browser, including earlier essential-only choices. Open cookie settings in the footer to review or change your choice; simply opening settings does not change it. This browser choice does not control server-side MCP activity analytics or product events. When configured, Google Analytics 4 uses cookies to measure website visits and pageviews, and Google Ads uses cookies to measure which of our ads led to visits and sign-ups. Both follow the same browser analytics choice, stay off when your browser sends Global Privacy Control, and are not used for ad personalization or remarketing. Task performance counting does not use cookies.

## Sensitive personal information

Identity verification involves sensitive personal information: your government ID and biometric data from your selfie. Persona collects and processes these under the consent it asks for and its own privacy policy and retention schedule; we do not receive them. We use the name on your ID only to verify your identity and the hash only to keep one ID from verifying more than one account, and we do not use sensitive personal information to infer characteristics about you.

## Background checks

Identity verification is not a background check. Toma does not run background checks or employment verification today and does not request Social Security numbers or dates of birth for background checks. Before we offer background checks, we will update this policy and provide the disclosures and authorizations required by the Fair Credit Reporting Act.

## Data retention

- Accounts, profiles, tasks, and applications: while your account is active. Account deletion removes its profile, photos, personal files and files on tasks you created, tasks and legacy drafts and recordings, applications, sessions, API keys, and agent access, and closes its open tasks. Stored-file and recording cleanup must finish before deletion completes; if it fails, deletion must be retried. A deletion you ask us for by email is completed within 30 days. We keep information only where we must to meet legal obligations, resolve disputes, or enforce agreements. Payment records are kept as tax and accounting law requires. Messages you sent to other members and comments may be kept without your name and photo.
- Unfinished sign-ups: accounts that don't finish the sign-up steps are deleted automatically once they are 7 days old, along with their profile, photos, and identity check records, and we ask Persona to delete their identity check. Deletion runs in the background, so it can take longer.
- Identity verification: we keep the name on your ID, the outcome, Persona's reference, and the hash of your ID number and birthdate while your account exists. Persona keeps the images and biometric data under its own retention schedule, which it states in its consent screen; when your account is deleted, we ask Persona to delete them.
- Personal files remain until their owner deletes them or deletes their account. Account deletion closes the personal file area and removes its bytes before deleting the account; failed cleanup must be retried. Private task files remain until an authorized participant deletes them or the poster deletes their account. Deleting a worker account does not delete files already provided to the task. Account deletion closes the poster’s file areas and removes their stored files first; if storage cleanup fails, deletion must be retried. No automatic file-retention deadline is set. Storage-provider versions and backups follow the deployment’s retention settings, and recipients may retain copies they already downloaded.
- Notification delivery records are kept for up to 30 days. Your email preferences remain until you delete your account. You can turn off these email alerts in your profile or through their unsubscribe link. The one-time registration welcome, account, and security emails are separate from optional alerts.
- Task source tracks, prior takes, uploaded evidence copies, retained answers and agent context, review questions, rendered revisions and excerpts, transcripts, privacy findings, captions, sampled frames, and thumbnails remain until account deletion; no automatic retention deadline is set. Discarding an unpublished task revokes access but does not immediately erase these stored objects. Removing evidence from a task or revoking a phone grant does not recall previously received files or immediately erase retained source objects. Legacy recordings, transcripts, and summaries remain until the poster deletes the recording or draft, or the account is deleted. Storage versions and backups follow the deployment's retention settings, and previously downloaded copies cannot be recalled.
- AI provider retention: Vercel distinguishes its Gateway retention policy from the policies of the model providers it routes to. Live transcription does not request Gateway zero-data-retention routing, so we cannot guarantee that its audio and transcript are deleted immediately by the provider. xAI's standard API policy describes 30-day retention without model training; written agreements and legal requirements can change retention. That standard policy is not a verified retention setting for Vercel's live-audio route. Recorded-audio transcription and text/image analysis request Gateway zero-data-retention routing.
- Sessions expire after 7 days of inactivity. API keys and agent access last until you revoke them or your account is deleted.
- Task performance: daily counts per task are kept for the life of the task. The daily keys used to count each viewer once are deleted after 2 days.
- Logs: for the period set by our hosting provider, typically no more than 30 days.
- PostHog analytics and session replay: for the retention periods configured in our PostHog project. Turning off browser analytics stops future collection; contact us to request deletion of previously collected data.

- [Vercel AI Gateway Data Retention](https://vercel.com/docs/ai-gateway/security-and-compliance/zdr)
- [xAI API Data Handling](https://docs.x.ai/developers/faq/security)
- [xAI Enterprise Data Retention](https://x.ai/legal/faq-enterprise)

## Data security

We protect information with encryption in transit (TLS), encryption at rest through our infrastructure providers, hashed API keys, access controls, rate limiting, and monitoring. No method of transmission or storage is completely secure, and we cannot guarantee absolute security. If you believe your account has been compromised, contact us immediately at support@toma.com.

## Your rights and choices

You can view and update your profile and photo in the Services at any time, discard unpublished tasks to revoke access, delete legacy recordings and task drafts you created, revoke API keys and agent access, and delete your account from your profile page (or have your agent do it for you).

California residents have rights under the California Consumer Privacy Act as amended by the California Privacy Rights Act (CCPA/CPRA): to know what personal information we collect, use, and disclose; to delete it; to correct it; to opt out of its sale or sharing (we do neither); to limit the use of sensitive personal information (we use it for identity verification and to store and transfer private files you choose to share); to data portability; and not to be discriminated against for exercising these rights. Residents of other states with privacy laws, such as Virginia, Colorado, Connecticut, and Utah, may have similar rights.

To make any other request, or if you can't sign in to delete your account, email support@toma.com with the subject "Privacy request" (or "Delete my data" for deletion) from the address on your account. We verify your identity before acting, acknowledge requests within 10 business days, and respond within 45 calendar days, which we may extend by up to 45 more days with notice. An authorized agent may submit a request with your signed written permission.

## Children

The Services are not intended for anyone under 18, and we do not knowingly collect personal information from children under 13. Members must be at least 18. If you believe a child has given us personal information, contact us and we will delete it.

## Geographic scope

Toma is operated from the United States, and our infrastructure and support are located there. Accounts are open only to people who live in and are authorized to work in the United States. We do not intentionally market to or knowingly collect personal information from residents of the European Economic Area or the United Kingdom.

## Changes to this policy

We will post changes on this page and update the date above. For material changes, we will also notify signed-in users by email or in the product. Continued use of the Services after changes take effect means you accept the updated policy.

## Contact us

Toma, Inc., 277 Carolina St., San Francisco, CA 94103. Email: support@toma.com.

- [Terms of service](https://www.toma.com/terms)

HTML: https://www.toma.com/privacy