A sandbox per thread
Each thread gets its own isolated sandbox running your real app stack — services, workers, databases — behind its own preview URL. Work in one thread can’t see another.
Enterprise & trust
Built to be let loose on the app you actually care about.
Giving an AI agent a real, complex product is a trust decision. Here is how Gentlybot earns it: isolated sandboxes, your own cloud and model accounts when you want them, encrypted secrets, opt-in autonomy, and a review trail for every change.
Isolation
Where the work happens.
Hard resource boundaries
Sandboxes run as isolated containers or lightweight microVMs (with their own kernel) depending on the node, with CPU, memory, and disk limits enforced per sandbox.
Code stays in GitHub
Gently doesn’t keep a copy of your repositories. Code is cloned on demand into the sandbox for a run through your GitHub App installation, which you can revoke at any time.
Bring your own
Your cloud, your models — when that’s the requirement.
Start on Gently’s managed capacity, or keep compute and model traffic inside your own accounts. The product works the same either way.
Gently cloud
The default: managed sandbox capacity with per-thread microVM isolation, no infrastructure to run.
Your own AWS
Run sandbox nodes in your own AWS account. Nodes carry no long-lived cloud credentials — artifacts move through short-lived presigned URLs only.
Your own models
Bring your own model accounts or private deployments across Anthropic, OpenAI, and open-source families. Admins choose exactly which models each pool may use.
Control
Autonomy you dial in, not out.
Roles, not free-for-all
Owner, admin, member, and viewer roles scope who can manage integrations, projects, and billing — and who simply asks, plans, or reviews.
Capabilities are opt-in
Autonomous behaviors — auto-fixing failing CI, addressing review comments, acting on security alerts — are individual toggles, all off by default, each with its own scoped GitHub permission.
A trail for everything
Every run lives in a thread: the prompt, the plan, the evidence, the PR. Changes reach production only through your existing review and CI — never around them.
Usage controls
Workspace-level budgets with admin alerts at 50/80/100%, pause-or-cap overage behavior, and per-org concurrency limits.
Secrets & data
What we store, and what we don’t.
- Integration tokens and model keys are encrypted at rest (AES-256-GCM), scoped to your workspace, and revocable.
- App secrets can stay in your 1Password vault — Gently resolves them into the sandbox at run time instead of storing them.
- Slack messages and Jira issues are read on demand through their APIs, not mirrored into Gently.
- Sandbox workspaces are ephemeral; idle sandboxes are reaped and per-thread artifacts expire on a short retention window.
Have a security questionnaire?
We’re onboarding teams carefully and answer security reviews directly. Request access, mention your requirements — private cloud, model accounts, data handling — and we’ll walk through the details with your team.