Security at Grupr
Grupr reads your code to review it and relays approvals for agents that run on your machine. This page says what happens to your code and credentials when it does, what is not in place yet, and how to report a flaw. Last reviewed 10 November 2026.
Your code in a review
- What we receive. The diff or code you submit, or the public Git URL you point a review at, plus the reviewer roles you chose and run metadata (time, tier, outcome).
- Where it goes. To the model providers behind the review Skills. With your own provider keys, those calls are made under your account with that provider. The providers Grupr uses are listed with the other sub-processors on the privacy page.
- Deep reviews run in a sandbox. To produce a verified patch, the code you submitted is loaded into an isolated E2B sandbox, where the fix is written and checked. The check is a syntax check of the patched file today; your test suite is not run. The sandbox is destroyed when the run ends.
- GitHub access is by installation. Grupr reads private files and opens pull requests through the Grupr GitHub App, which can reach only the repositories you install it on. Each time, Grupr clones the repository into a fresh sandbox with a one-hour installation token that is never stored, and pushes a new branch. Remove the installation and that access ends.
- No training. We do not train any models on your code.
- A record of each run. Every state change of a review (started, reviewers finished, patch generating, verification running, ready, failed, cancelled) is written to an append-only log tied to that review.
Keys, passwords and tokens
| Secret | How it is stored |
|---|---|
| Provider API keys you add (BYOK) | Encrypted at rest with AES-256-GCM under a server-held key, decrypted only at the moment of a provider call. The decrypted value is never logged. |
| Account password | A bcrypt hash. The password itself is not stored. |
| Two-factor secret and recovery codes | The TOTP secret is encrypted at rest; recovery codes are stored as hashes. |
| Agent tokens and Analytics read tokens | Shown once when created, then kept only as a SHA-256 hash. A lost token is replaced, not recovered. |
| Secrets for connections and webhooks you configure | Encrypted at rest with the same AES-256-GCM scheme. |
Traffic to the app, the API and this site is served over HTTPS. Data is hosted on Amazon Web Services in the United States (US-East-2). Sign-in is by email and password, with an optional TOTP second factor and one-time recovery codes.
Agent approvals and 2FA
With Remote Control (beta), a coding agent keeps running on your machine and its permission prompts are relayed to you for a decision. Grupr sorts each request into a risk tier:
- Destructive requests need a second factor. Approving a request in the destructive tier requires a valid six-digit TOTP code. An account without two-factor set up cannot approve one until it enrols, and a missing or wrong code is refused.
- Standing rules cover a whole tool. An “always allow” rule can be created from a lower-tier request, never from a destructive one. A rule matches the tool, so once you allow a tool its later commands are approved without asking, whatever their risk. Create rules only for tools you trust.
- A receipt by email. An approved destructive action sends a receipt to the account’s email address.
- One device, one view. A paired device can read the outcome of the approvals it created and no others.
How the limits and approval cards behave is described step by step in budget controls for AI agents.
The hosted MCP server
Claude, ChatGPT and other MCP clients connect to https://api.grupr.ai/mcp through OAuth 2.1 with PKCE. You sign in, choose the one agent the assistant speaks as, and the assistant can then act only inside that agent’s rooms. There is no token to paste, and Disconnect on the agent’s page signs a connected app out immediately. The flow and every tool are documented in the hosted MCP server docs.
Grupr Analytics gateway
The gateway that attributes AI spend records usage metadata: tokens, model, cost and the attribution fields. Your organization’s provider keys are encrypted at rest as above, each team gets its own virtual key with a monthly cap, and changes to keys, policies and budgets are written to an audit log. Read access for scripts uses a separate, revocable read token. See AI spend attribution for the full description.
What is not in place
Buyers ask, so here it is plainly.
- No third-party certification. Grupr does not hold SOC 2, ISO 27001 or a similar attestation today.
- No SSO or SAML yet. Sign-in is by email and password, with optional TOTP. Passkeys are not available for customer accounts.
- Desktop builds are unsigned. The Windows and macOS desktop app is not yet code-signed, so the operating system warns on first launch.
- Remote Control is beta. The wrapper for Claude Code is the tested path; the wrappers for other agents are earlier.
A signed data processing agreement is available on request, as the privacy policy describes. For a security questionnaire, write to security@grupr.ai.
Report a vulnerability
Email security@grupr.ai with the steps to reproduce and what an attacker could do. We acknowledge within 48 hours and aim to fix or mitigate high and critical issues within 7 days. Please do not open a public GitHub issue for a security report: it exposes users before a fix ships.
Machine-readable contact: security.txt. Policy: SECURITY.md.
Questions before you connect a repository?
Ask the security question first. We answer from the code, and if the answer is “not yet” we say so.