At accessiBe, we understand that granting an application access to your proprietary source code requires an immense amount of trust. We designed accessFlow Code Agent from the ground up with a “Privacy-First, Stateless IP” architecture to ensure that your intellectual property remains completely secure and compliant with enterprise security standards (SOC2, GDPR, ISO 27001).
Here is a transparent breakdown of exactly how we handle your data, your source code, and your GitHub platform access.
Zero source code retention (stateless IP)
We never store, persist, or back up any of your source code.
When you open a Pull Request, accessFlow Code Agent temporarily streams the Git Diff into our isolated, ephemeral container memory to perform the AI-driven accessibility audit.
Once the audit is complete, the originalCode and the generated suggestedFix snippets are immediately pushed directly to your GitHub repository as PR Review Comments. The instant that network request completes, the code is wiped from our server’s RAM via standard garbage collection.
What do we store?
We strictly store Metadata (the “GPS coordinates” of a bug). Our database schema only records:
- The repository name and PR number.
- The ID of the accessibility rule that failed (e.g.,
image-alt). - The severity of the issue (e.g.,
critical). - The file path (e.g.,
src/components/Button.tsx). - The line number.
- The lifecycle status of the issue (
OPEN,RESOLVED,DISMISSED).
By completely decoupling the metadata from the actual source code, it is mathematically impossible for an attacker to extract your proprietary algorithms or business logic from our databases. GitHub remains the absolute single source of truth for your IP.
LLM processing & AI privacy
We utilize Enterprise-grade Large Language Models (LLMs) to perform the accessibility audits.
- No training on your data: Our LLM providers (e.g., Google Vertex AI / Gemini) are contractually bound under Enterprise Service Level Agreements (SLAs). Your source code and PR diffs are never used to train, fine-tune, or improve public AI models.
- Ephemeral processing: Your code is processed in secure, ephemeral inference environments and is discarded immediately after the review is generated.
- Our models, our cost: accessFlow Code Agent runs on accessiBe's own enterprise LLM accounts. We never ask for, connect to, or consume your AI provider API keys, Copilot seats, or any other token or credit balance of yours. All AI processing is provided and paid for by accessiBe.
- Continuous improvement via feedback: While your code is never used to train public models, our engine continuously learns from your team's direct interactions. By analyzing explicit feedback (like thumbs up/down reactions) and resolution states (whether an issue was dismissed or fixed), we dynamically refine our internal evaluation rules and prompts to adapt to your specific coding patterns over time.
GitHub permissions & access
The accessFlow Code Agent GitHub App operates with the absolute minimum privileges required to function (Principle of Least Privilege).
PR scan permissions
- Pull requests (read & write): Reads code changes in your pull requests (PRs) and posts inline accessibility suggestions and comments.
-
Contents (read-only): Fetches the full content of modified files and resolves repository configuration files (like .
codeagent.yml). - Checks (read & write): Creates and updates Check Run summaries on your PRs.
- Issues (read & write): Powers the interactive chat interface. GitHub treats top-level PR conversation comments as issues, which allows accessFlow Code Agent to read and respond to slash commands (for example, /codeAgent help).
Repo scan permissions
- Pull requests (read & write): Reads code changes in your pull requests (PRs) and posts inline accessibility suggestions and comments.
-
Contents (read & write): Fetches the full content of modified files and resolves repository configuration files (like .
codeagent.yml). - Checks (read & write): Creates and updates Check Run summaries on your PRs.
- Issues (read & write): Powers the interactive chat interface. GitHub treats top-level PR conversation comments as issues, which allows accessFlow Code Agent to read and respond to slash commands (for example, /codeAgent help).
- Events (push) - Git push to a repository - enables pushing permission to the selected repos.
Updates to Code Agent's requested Github permissions take effect only after an organization admin reviews and accepts Code Agent’s updated permission request in GitHub, and selects Accept new permissions.
Token security
- We do not ask for or store Personal Access Tokens (PATs).
- accessFlow Code Agent uses dynamic, short-lived installation tokens generated securely via GitHub App private key cryptography. These tokens expire automatically and are rotated continuously, ensuring that even if a token were intercepted, it would quickly become useless.
4. Telemetry & analytics
To help your organization track its ROI and accessibility health, we collect high-level telemetry data. This data is strictly aggregated and contains absolutely no sensitive information or PII.
Telemetry events logged:
-
AUDIT_STARTED/COMPLETED/FAILED -
ISSUE_CREATED/ISSUE_AUTO_RESOLVED/ISSUE_MANUALLY_DISMISSED -
REACTION_ADDED(when a developer reacts with a 👍/👎 to an AI comment). -
LLM_USAGE_METRICS(Token counts and latency for billing/performance tracking).
Telemetry events only reference the UUIDs of the metadata rows mentioned in Section 1 (rows in the PR along with the file name). They never log the actual code.
Right to erasure & GDPR
We only store metadata and relational IDs.
For more information about how we handle your data and respect your rights, we encourage you to view our Privacy Notice.
Uninstalling the accessFlow Code Agent does not automatically delete your stored data. If you want us to delete historical telemetry and metadata associated with your account, you must submit a deletion request to our support team.
Once we receive and process your request, we will delete the applicable data in accordance with our data retention policies and relevant data protection requirements.
If your Infosec or Compliance teams have any further questions or require a formal security questionnaire, please contact our enterprise support team.