Last revised: . Policy version: 2026-10-31.
This policy describes how Dharma AI Labs Inc. ("Dharma AI", "we", or "us") processes information through our website and services, including Dharma Agent Fabric, its APIs, local agent integrations, and the Dharma Agent Fabric ChatGPT/MCP app. Your organization may also set policies governing its workspaces, agents, evidence, and learning features.
We process information you provide, information sent by connected clients and agents, and records created when you use the service. Depending on the feature and the organization's configuration, this includes:
Do not assume code, prompts, messages, attachments, or evidence are anonymous. Their contents and associated identifiers may identify you, your organization, or other people.
Agent Fabric receives data in two ways. First, an MCP client such as ChatGPT sends tool calls to our hosted MCP server at mcp.dharma-ai.io. Each call is authenticated with OAuth, checked against your active organization membership and handled by the same organization APIs as the Dharma portal. Second, the Agent Fabric relay on an enrolled device opens outbound connections only. It uploads the records your organization's policies allow, and it receives tasks, evidence requests and skills.
Every tool needs an organization ID. Every tool that changes something, and every tool that expands stored evidence, needs a confirmation field set to true. Every change also needs an idempotency key.
| What we collect |
|
|---|---|
| Purpose | Authenticate you, check your active organization membership on every tool call, and show which identity is connected. |
| Recipients | Clerk (sign-in and OAuth). Supabase (membership records). Resend (when an account is created, a notification with the new user’s name and email address is emailed to a Dharma staff mailbox). The MCP client you use receives your user ID, email address and membership state when you call fabric_get_identity. |
| Retention | Kept while your account or organization membership exists. Account and membership records are deleted 30 days after an account is closed or a member is removed, except billing and audit records, which follow the retention for usage, billing, audit and security records below. |
| Your controls | Disconnect the Dharma Agent Fabric app in your ChatGPT settings to stop ChatGPT from calling our tools. Organization administrators can remove members. Contact us to close your account or revoke the OAuth grant on our side. |
| What we collect |
|
|---|---|
| Purpose | Enroll devices, route work to the right endpoint, enforce which endpoint may do what, and show your organization’s agents. |
| Recipients | Supabase. GitHub, for repositories you connect. Google Cloud, for managed runtime endpoints and for customer-owned Vertex AI projects reached through short-lived workload identity federation. |
| Retention | Kept while your organization uses Agent Fabric. Records of revoked devices, workspaces and endpoints are kept for 12 months as audit records, then deleted. |
| Your controls | Revoke a device (this stops new relay work at once), pause or revoke a workspace, unbind an endpoint, or uninstall the GitHub App. |
| What we collect |
|
|---|---|
| Purpose | Show agent activity, run deterministic checks and failure analysis, and decide whether a deeper review is recommended. |
| Recipients | Supabase. Members and MCP clients of your organization that are authorized to read trajectories. |
| Retention | A retention job runs every 10 minutes. It replaces capsule content with a tombstone 90 days after upload. The tombstone keeps the capsule hash, its byte count and its disclosure class. An analysis in progress or an open evidence request can delay this. |
| Your controls | Organization evidence policy, device revocation and deletion requests. |
| What we collect |
|
|---|---|
| Purpose | Continuous evaluation of your organization’s agents, which is the only purpose this consent allows, and the purpose stated in each evidence request. |
| Recipients | Supabase. Google Cloud Vertex AI (Gemini models in Dharma’s Google Cloud project) for semantic analysis and learning reviews. Members and MCP clients of your organization that are authorized to read the evidence. |
| Retention | Synchronized content expires after the retention period the administrator chose, from 1 to 365 days (the portal suggests 30). Evaluation content derived from it expires at the same time. Responses to 24-hour evidence requests are removed 24 hours after the request completes. Responses kept under organization policy expire with that policy. |
| Your controls | Only an administrator can turn this on, and must confirm it and choose the retention period and byte caps. Withdrawing consent at once replaces the synchronized content and its derived evaluation content with tombstones. If a bounded analysis is using the content at that moment, withdrawal is refused until it finishes. Each evidence request needs an explicit confirmation. |
| What we collect |
|
|---|---|
| Purpose | Build, evaluate and deliver skills and remediation packages for that repository. |
| Recipients | Supabase. Google Cloud Vertex AI for reviews. GitHub, when a package is delivered to your repository as a pull request or check. |
| Retention | Uploaded content is stored as trajectory evidence and follows the retention rules above. |
| Your controls | An administrator can turn the organization source policy off so that new repositories are no longer authorized automatically. For each repository, an administrator can later narrow the approved paths, change the byte caps, set an expiry date or revoke the authorization. |
| What we collect |
|
|---|---|
| Purpose | Run the work you request on your own enrolled endpoint, and pass bounded state between your agents. |
| Recipients | Supabase. Google Cloud KMS, which receives the bytes of each task envelope in order to sign it. The agent provider on your device (for example Codex or Claude Code) then processes the instructions under that provider’s own terms. |
| Retention | Task and handoff records are kept for the life of your organization unless you ask us to delete them. |
| Your controls | Every dispatch needs an explicit confirmation. Revoking a device stops it from receiving tasks. |
| What we collect |
|
|---|---|
| Purpose | Answer questions about your organization and record proposed actions for you to approve or reject. |
| Recipients | Supabase. Google Cloud: Dharma’s managed agent runtime on Google Cloud and a Gemini model receive your message, the recent conversation, a bounded summary of your organization and any images, for that turn. |
| Retention | Images are passed to the runtime for the turn. For most organizations Dharma stores only each image’s file name, type and hash. Organizations enrolled in automatic learning are different. Dharma staff enroll an organization, and capture also needs that organization’s active learning policy and content-synchronization consent. For an enrolled organization, the full request for each turn is kept as learning input, including your message, the conversation and organization context sent with it and the image bytes. It follows the learning retention below. Conversation messages and events are kept for the life of your organization unless you ask us to delete them. |
| Your controls | Creating a session and sending a message each need an explicit confirmation. A proposal does nothing until an organization owner or administrator approves it in the portal, or runs its separate execution tool and confirms it. |
| What we collect |
|
|---|---|
| Purpose | Run the agent work your organization starts, measure and compare your agents, and bill the credits used. |
| Recipients | Supabase. Google Cloud: Dharma’s managed agent runtime and Gemini models in Dharma’s Google Cloud project, or a customer-owned Google Cloud project (BYOK) that your organization bound as the endpoint, through short-lived workload identity federation. If your organization is enrolled in automatic learning, the run input is also kept as learning input (see the control-agent and learning categories). |
| Retention | Evaluation tasks built from synchronized content expire with that content: the retention job replaces their task content, results and run-event messages with tombstones. Managed run records, run events, control-agent proposals and evaluation campaigns that are not built from synchronized content are kept for the life of your organization unless you ask us to delete them. |
| Your controls | There are two ways to start a run or campaign. First, directly from the portal or the API: any active organization owner, administrator or developer can start one, and so can an API token with the agents:run scope (runs) or the evals:run scope (campaigns). An agents:run token can also start runs through the Gemini-compatible generateContent API. Owners, administrators and developers can create API tokens. This way needs no proposal and no administrator approval. Second, through the control agent: the run or campaign starts from a stored proposal, and only when an organization owner or administrator approves it in the portal or runs fabric_control_execute_managed_run or fabric_control_execute_evaluation_campaign. Those two MCP tools also need an explicit confirmation and the server’s hash of the stored arguments, and a proposal runs at most once. Either way, the organization’s managed runtime must be activated, and each run reserves credits before it starts (credit preauthorization). A campaign must also fit within the runtime spend ceiling. A direct run request needs an Idempotency-Key, and retrying with the same key returns the same run instead of starting another. The Gemini-compatible API accepts the key but does not require it. A direct campaign request has no such key, so each request creates a new campaign. |
| What we collect |
|
|---|---|
| Purpose | Evaluate and improve your organization’s own agents. Dharma’s rules do not allow this data to be used for another customer or to train models. |
| Recipients | Supabase. Google Cloud Vertex AI for analysis and reviews. Google Cloud KMS signs skill bundles and receipts. GitHub, where an approved remediation is delivered as a pull request, check or merge (fabric_remediation_merge_pr merges only the exact approved pull request). Your endpoints receive released skills. |
| Retention | Learning copies of content expire no later than the consent and the source content they came from. When a learning window’s source content expires or consent is withdrawn, the model responses, failure observations and learned skill bundles derived from it are replaced with tombstones or retired. Signed release history, hashes and receipts are kept. |
| Your controls | Pause or revise a learning policy, reject a proposal, roll back a rollout. Approvals and confirmations are required at each gate. |
| What we collect |
|
|---|---|
| Purpose | Bill and account for usage, prevent duplicate actions, investigate incidents and keep the service secure. |
| Recipients | Supabase. Stripe (payments). Vercel (hosting and request logs). |
| Retention | Billing records are kept for 7 years to meet accounting obligations. Audit and idempotency records are kept for 12 months. Hosting logs are kept for the period our hosting provider, Vercel, retains them. |
| Your controls | Contact us for a copy of your records. Some records must be kept to meet legal and accounting obligations. |
The hosted MCP server offers 65 tools, and ChatGPT and other clients can call any of them. The client sends the tool arguments listed below. We send back a JSON result for your organization, or an error code. A read tool can show existing organization data to the client that calls it.
| Tools | fabric_get_identity, fabric_list_devices, fabric_list_agents, fabric_list_repository_agents, fabric_get_instructions, fabric_list_workspaces |
|---|---|
| Inputs sent | Organization ID. |
| Results returned | Your user ID, email address, membership role and binding state, and the connected OAuth client and scopes; device names, platforms and status; agents and endpoint bindings; repository agents; setup instructions; workspaces and provider capabilities. No local paths or credentials. |
| Tools | fabric_create_agent, fabric_bind_local_endpoint, fabric_bind_runtime_endpoint |
|---|---|
| Inputs sent | Agent key, name and description; agent, workspace, provider and runtime binding IDs; priority. |
| Results returned | The created agent or endpoint binding. |
| Tools | fabric_list_trajectories, fabric_get_trajectory_evidence, fabric_request_trajectory_evidence |
|---|---|
| Inputs sent | Trajectory ID and a result limit; for an evidence request, its purpose, content selectors (content hashes, ranges, reasons), a byte cap, a retention class and an authority decision ID. |
| Results returned | Reduced trajectory summaries; for one trajectory, its stored revision capsules, which include synchronized content if your organization turned it on; the evidence request record. |
| Tools | fabric_list_tasks, fabric_list_a2a, fabric_dispatch_task, fabric_dispatch_a2a |
|---|---|
| Inputs sent | Task instructions, target endpoint and workspace, allowed paths, commands, network mode and domains, time limits, required artifacts and skills; for a handoff, the source task, target endpoint, state envelope and evidence references. |
| Results returned | Task state and signed envelope hashes; handoff records; the dispatched task or handoff. |
| Tools | fabric_list_evals, fabric_run_analysis, fabric_list_failure_atlas, fabric_list_remediations, fabric_remediation_stage_evaluation, fabric_remediation_run_backtest, fabric_remediation_retry_backtest, fabric_remediation_link_backtest, fabric_remediation_approve, fabric_remediation_merge_pr, fabric_remediation_release, fabric_remediation_retry_release, fabric_remediation_verify_release, fabric_remediation_expand, fabric_remediation_rollback |
|---|---|
| Inputs sent | Analysis scope (organization, agents or endpoints) and size; remediation target ID; endpoint ID, trajectory ID lists or campaign ID where the step needs them; whether to set up automatic updates. |
| Results returned | Analysis windows and results; failure families with recurrence and lineage; remediation campaigns, gates, approval, pull request, release and rollout state and receipts. |
| Tools | fabric_list_skills, fabric_release_skill, fabric_rollout_start, fabric_rollout_expand, fabric_rollout_rollback |
|---|---|
| Inputs sent | Skill revision, skill ID, risk class, activation policy, rollback bundle, approval ID and target selectors; bundle ID and canary percentage. |
| Results returned | Signed bundles, rollout state and installation receipts. |
| Tools | fabric_list_learning_policies, fabric_get_learning_policy, fabric_list_learning_windows, fabric_create_learning_policy, fabric_update_learning_policy, fabric_submit_calibration_references |
|---|---|
| Inputs sent | Policy ID; the full consent-bound configuration with spending caps; expected revision; a signed calibration manifest (up to 512 KB) and its signature. |
| Results returned | Policy configuration, evidence coverage, costs, candidate and rollout progress, learning windows and status. Raw trajectories are not returned. |
| Tools | fabric_list_usage |
|---|---|
| Inputs sent | Organization ID. |
| Results returned | Settled credits and provider usage for your organization. |
| Tools | fabric_control_agent_list_sessions, fabric_control_agent_create_session, fabric_control_agent_send_message, fabric_control_agent_list_events, fabric_control_read_org_status, fabric_control_read_agents, fabric_control_read_trajectories, fabric_control_read_failures, fabric_control_read_remediations, fabric_control_read_usage, fabric_control_read_runs, fabric_control_read_evaluations, fabric_control_execute_analysis, fabric_control_execute_managed_run, fabric_control_execute_evaluation_campaign, fabric_control_execute_task, fabric_control_execute_a2a, fabric_control_execute_remediation_approval, fabric_control_execute_rollout_start, fabric_control_execute_rollout_expand, fabric_control_execute_rollout_rollback, fabric_control_reject_proposal |
|---|---|
| Inputs sent | Session ID and title; your message and up to two images; a result limit; for an execution or rejection, the proposal ID and the server’s hash of its stored arguments, and for a managed run or evaluation campaign the matching work details (such as the prompt). |
| Results returned | Sessions, messages, run, response and approval events; bounded organization summaries (status, agents, trajectories, failures, remediations, usage, runs, evaluations); the result of an executed or rejected proposal. |
ChatGPT (OpenAI) and any other client you choose receive your tool arguments and our tool results, and process them under their own terms and privacy policies. ChatGPT has its own controls for conversation retention, memory and model improvement. Dharma's organization policies do not control those settings. Before you confirm an action, review your client's settings, the tool arguments and the information you choose to retrieve or send.
We use this information to authenticate users and enforce organization access; enroll and connect agents and endpoints; carry out requested tasks and agent handoffs; provide control-agent conversations and approvals; deliver evidence, evaluations, failure analysis and remediation; operate enabled learning, calibration and skill testing or rollout; provide support; measure usage and administer billing; maintain service reliability and security; and fulfill our contractual and legal obligations.
Configured learning uses relevant evidence and evaluation or calibration results through model inference to assess and improve the configured agents, instructions, rubrics and skills. Dharma's rules limit use of customer content and content-derived artifacts to that customer's requested operations and configured learning. These rules prohibit use across customers or for model training and also apply to derived evaluations, calibration references, skills, failure patterns and provider receipts. Service processing and learning are separate from optional website analytics. An analytics opt-out does not disable an agent task or learning policy.
Information is available to the members, administrators, agents and endpoints authorized for the applicable organization and operation, and to the client receiving a requested tool result. A task, handoff or rollout may send content to the selected agent or provider endpoint. Repository artifacts and release information may be delivered to the GitHub repositories you connect for that operation.
Dharma uses the following service providers (subprocessors) to run Agent Fabric and the ChatGPT app. Which of them handle a record depends on the feature and your organization's settings, as shown in section 3.1. Other Dharma products may use additional providers, which are listed in their own terms.
| Provider | What it does | Data it handles |
|---|---|---|
| Vercel Inc. | Hosts the website, the APIs, the MCP server at mcp.dharma-ai.io and scheduled jobs. | All requests and responses in transit; request logs, including IP address and user agent. |
| Supabase Inc. | Main database for Agent Fabric records. | Every stored category in the tables above. |
| Clerk Inc. | Sign-in, organizations and OAuth for ChatGPT and other MCP clients. | Name, email address, user and organization IDs, sessions and OAuth grants. |
| Google LLC (Google Cloud) | Vertex AI Gemini models for analysis, learning reviews and the control agent; the managed agent runtime; Cloud KMS signing; identity federation into customer-owned projects. | The content each feature needs (synchronized content, repository content, control-agent messages and images, learning material) and the bytes of signed task, skill and receipt payloads. |
| GitHub Inc. | The Dharma GitHub App for repositories you connect: pull requests, checks, merges and identity linking. | Repository identifiers, skill and remediation files, pull request and check details, and your GitHub user identity if you link it. |
| Stripe Inc. | Payments for subscriptions and credits. | Billing contact and payment details, which Stripe collects directly, and transaction references. |
| Resend | Transactional email, including a new-account notice to Dharma staff and operational alerts. | Recipient address and the content of those emails, which for the new-account notice includes the new user’s name and email address. |
If you accept analytics, Google Analytics also receives website data, as described in section 8. It does not receive Agent Fabric service data.
Data sales assessments. For a data sales assessment, Dharma also uses these providers. OpenAI provides the models that analyze your submission. Cloudflare carries the network traffic to and from the processing server. GitHub holds the job files and the finished report in a private Dharma repository. Google Cloud runs the processing server.
Other recipients are ones you choose. The MCP client you connect, such as ChatGPT, receives tool results. The agent providers on your own devices, such as Codex, Claude Code, Agy or Hermes, process tasks and skills under your own accounts with them. A customer-owned Google Cloud project you bind as a bring-your-own-provider endpoint receives the content of the runs you send to it. Stripe collects payment details directly. Our billing records hold transaction references, usage and purchase information.
We process data primarily in the United States. Vertex AI requests may use Google's global endpoint. We will give customers 30 days' notice before we add a new subprocessor.
We may also disclose information when required by law, to protect rights and service security, or as part of a business transfer subject to applicable protections. Hashed, summarized or derived records are not automatically anonymous.
The specific Agent Fabric retention rules are listed for each category in section 3.1. In brief: default-mode trajectory capsules become tombstones 90 days after upload. Synchronized content expires after the period your administrator chose (1 to 365 days), or at once when consent is withdrawn. Responses to 24-hour evidence requests are removed 24 hours after the request completes. Learning material expires with the content it came from.
Retention differs by record type and by the policy applicable to the organization, evidence grant or learning operation. Local evidence storage and server storage have separate lifecycles. Completing a task, disconnecting a client, pausing learning or withdrawing future collection does not by itself delete every existing record.
Content-expiry settings cover particular record types and do not establish that all related records or backups have been erased. Deletion coverage and timing differ across stores and providers. Active analysis, evidence requests or content leases can delay deletion of expired content. Removing content can leave a tombstone, content hash, lineage, status, validation or audit metadata. Evaluation records, calibration references, learned or released skills, task and conversation records, billing records and legal acceptance records can have different retention needs from the original content.
We use the account or service relationship, the purpose of the record, active work, security and audit needs, accounting and legal obligations, and applicable organization or contractual policy to determine retention. There is no single deletion deadline that applies to all service records. Backup copies and data held by a selected client or provider require separate consideration.
Contact us to request the applicable retention information or deletion of your data, including associated content and derived records. We will assess the request, the relevant organization authority, applicable obligations and any records that must be retained. Do not rely on a local vault default or an advertised content-expiry setting as a guarantee that every server record or backup has been erased.
Use your organization's available evidence and disclosure policies to limit what agents send. Where offered, review an upload or evidence preview, select bounded evidence, and pause or withdraw content synchronization. Organization administrators can manage membership, permissions and endpoint access. Learning policies can be paused or revised. Designated sensitive or mutating MCP calls require a confirmation field set to true.
Stopping future collection or revoking access is different from deleting data already collected. CEO Robert Cirino is the accountable owner for privacy requests. To request access, correction, export or deletion, or to ask about an existing record, contact support@dharma-ai.io. Identify the relevant organization and service without sending passwords, provider keys or other secrets. We may need to verify your identity and authority for organization data.
If you accept analytics, we use Google Analytics to understand page visits and named product-journey actions. Our event code adds a shortened page path (its first URL path segment), a bounded action name for journey events, the application surface and deployment environment. Google's tag may also collect the current page URL (including paths and query strings), page title, referrer, browser/device information and cookie identifiers under Google's policies and settings. We do not intentionally add task or evaluation payloads or account profile fields as custom event parameters. These analytics records are separate from Agent Fabric service processing.
Google Analytics does not load before you accept. Your choice is stored in the dharma_analytics_consent cookie for up to one year; you can reset the choice by deleting that cookie in your browser and reloading the page.
We use access controls and other measures designed to protect information from unauthorized access, use, alteration and disclosure. You are responsible for protecting your account and the credentials and permissions of your connected integrations. No security measure eliminates every risk.
Dharma AI is not a HIPAA-compliant platform for storing Protected Health Information. Do not submit patient information or PHI through our services, or include passwords, access tokens or other secrets in task instructions, messages or evidence.
We publish changes on this page and identify the actual revision date and policy version above. Material changes to our treatment of personal information will be announced through the website. A new published version does not rewrite historical legal acceptance records.
For privacy questions or requests, email support@dharma-ai.io. Dharma AI Labs Inc. is responsible for this processing.