You pick the repositoriesVocxAI is a GitHub App. You choose all repositories or specific ones on GitHub's own install screen, and only those are ever visible to us.
Three permissionsContents, Pull requests and Metadata. Nothing else: no Actions, no admin, no secrets, no organization settings.
Nothing merges itselfAgents work on their own branch and open a pull request. A person reviews it and decides. Our code never merges.
Short-lived credentialsRepository access uses GitHub tokens minted on demand that expire within the hour. We don't store them.
No training on your dataWe don't train or fine-tune models on your code or feedback, and we call model providers through their commercial APIs.
Revoke any timeDisconnect in VocxAI, and uninstall the app on GitHub for a hard stop. Access ends the moment you remove it.
What VocxAI can access on GitHub
VocxAI connects as a GitHub App installation. It does not use your GitHub password or a personal access token. The permissions below are the ones the app requests, and GitHub publishes them on the VocxAI app page, so you can check them against this list yourself.
| Permission | Level | What we use it for |
|---|
| Contents | Read & write | Read the repository tree, README, build and config files, and code-search results, so PRDs and feasibility checks reflect your real code. Create a branch and commit the changes a build agent makes. |
| Pull requests | Read & write | Open a pull request for each build, so the change arrives for human review. |
| Metadata | Read | List which repositories the installation covers. GitHub requires this of every app. |
What we don't ask for
- Actions and workflow files. Without the workflows permission, GitHub itself refuses any push that changes
.github/workflows, so agents can't alter your CI/CD pipelines.
- Administration, secrets and deployments. No repository settings, no Actions secrets or variables, no deploy keys, no deployments.
- Organization and member data. No members, teams or organization settings. Issues, packages and webhooks are also outside the app's permissions.
- GitHub events. VocxAI doesn't subscribe to webhooks from your repositories.
You choose which repositories
On GitHub's install screen you grant access to all repositories or to specific ones, and you can change that any time in your GitHub settings. After installing, VocxAI asks you to select which of the granted repositories it should actually analyze day to day. Nothing outside the granted set is visible to us.
About write access
Contents: write is the permission that lets an agent push a branch, and GitHub has no narrower one. It is technically enough to write to other branches in repositories you've granted. Our agents are set up to work on their own branch (feature/… by default) and open a pull request, and VocxAI never merges. To make that a guarantee instead of a convention, turn on branch protection with required pull-request reviews on your default branch. GitHub enforces it for every app and every person.
How credentials are handled
- No long-lived GitHub secrets. We store the installation ID, the account name and the list of repositories the install covers. We never hold your GitHub password or a personal access token.
- Tokens are minted on demand. When VocxAI needs to reach a repository, it asks GitHub for an installation token using the app's private key. GitHub expires these tokens within an hour. We don't save them. The private key itself lives in AWS Secrets Manager.
- The optional organization picker is read-only and one-time. If you belong to several organizations, an optional step signs you in to GitHub with the
read:org scope purely to list them, so the install screen can open on the right one. That token is used once and discarded.
- Other connected tools. Credentials for tools like your support desk or Slack are encrypted with an AWS KMS key before they are stored.
Where your code goes
Your code is not processed inside a black box, so here is each place it goes and why.
- Build agents. Agents run in sandboxed sessions on Anthropic's Claude Managed Agents platform. For each job, a short-lived token checks out only the repositories in scope, and the agent works on its own branch. When the job ends, we archive the session and delete the files we uploaded to it.
- Repository analysis. To build the repo summary, VocxAI reads the README and build and configuration files and asks Claude (through Anthropic's API) to describe the project's purpose, tech stack and structure.
- Feasibility checks and codebase questions. Code-search results from your granted repositories are passed to Claude so it can judge whether a capability already exists or what a change would touch.
- Customer feedback. The feedback you connect is embedded through OpenAI's embeddings API so similar requests can be clustered. That is feedback text, not source code.
VocxAI does not keep a mirror or clone of your repositories on its own servers. The finished code lives in your GitHub repository, as commits and pull requests.
What we store
- Connection details. The GitHub installation ID, account name and repository list.
- Derived repository notes. Summaries generated from your code: purpose, architecture, tech stack, patterns, entry points, known build commands and file paths.
- AI-written work products. PRDs, architecture decisions, specs and tasks, which reference the files and areas of your code they concern.
- Agent run logs. The streamed output of each build job, so you can see what an agent did. Because agents read and edit files, these logs can contain excerpts of your code.
- The customer feedback you connect. As described in our Privacy Policy, we process it on your behalf, as your data processor.
Data stays in your workspace until you ask us to remove it. There is no self-serve delete button yet: to have your workspace data deleted, ask us through the contact page.
AI providers and sub-processors
VocxAI does not train or fine-tune any model on your data. We use model providers through their commercial APIs, and under those providers' current API terms, content sent through the API is not used to train their models. These are the third parties that handle data on our behalf:
| Provider | Purpose | Data involved |
|---|
| Amazon Web Services | Hosting, database, key management and secrets. US West (Oregon). | Everything the platform stores |
| Anthropic | Claude models and Managed Agents: build agents, analysis, PRDs and chat | Prompts built from your feedback themes and the repository content described above |
| OpenAI | Text embeddings for clustering feedback | Customer feedback and idea text, not source code |
| Stripe | Billing | Billing contact and payment details, handled by Stripe's hosted checkout |
| Resend | Transactional email such as team invites and password resets | Recipient email address and the message |
| Slack | Internal alerts to the VocxAI team | Workspace name, user email, plan and connector name. No customer content or code. |
| Google Tag Manager | Analytics on this marketing site only | Standard site usage data |
How we protect it
- Encrypted in transit. The app is served over HTTPS only, with no plaintext HTTP listener, and connections to our database require TLS.
- Secrets encrypted with KMS. Connector credentials and other app-managed secrets are encrypted with an AWS KMS key, which rotates automatically, before they are written to the database. Platform secrets are held in AWS Secrets Manager.
- A private database. The database sits in private network subnets with no public access, reachable only from the application.
- Human approval before anything is built. A person approves the work before an agent starts, and reviews the pull request before anything ships. See human approval gates.
- Roles that limit who can send work to build. Workspaces have Owner, Admin, Developer and Product roles. Only Owners, Admins and Developers can send work to build, only Owners and Admins can manage billing or people, and these limits are enforced on the server, not just hidden in the interface.
- Access ends when it should. Removing someone from a workspace signs them out, and resetting a password ends every session that predates it.
- Passwords are hashed. We store salted password hashes, never the passwords themselves.
Staying in control and revoking access
You can end VocxAI's access to your code at any time, without asking us:
- Narrow it. In GitHub, change which repositories the app is installed on. The change takes effect immediately.
- Disconnect in VocxAI. In Settings → Connectors, disconnect GitHub. This removes our record of the installation.
- Uninstall on GitHub for a hard stop. Disconnecting in VocxAI does not uninstall the app from your organization. To remove all access, also remove the VocxAI app under your organization's (or account's) installed GitHub Apps. Once it's removed, no token can be minted for your repositories.
Everything VocxAI's agents produce arrives as an ordinary pull request. You can review it, ask for changes or close it like any other, and your existing branch protections and CI apply.
Reporting a security issue
If you believe you've found a vulnerability, please tell us through the contact page and put "Security" in your message. Please give us a reasonable chance to fix it before you share it publicly. We'll acknowledge your report and keep you informed.
Private beta
VocxAI is in private beta, and this page describes our practices in general terms. If your organization has a signed agreement with us that covers data handling, that agreement takes precedence. Questions about a security review or a questionnaire? Reach out through the contact page.
Related