Security & Trust
How Buda protects your data and isolates agent execution — so private files, internal SOPs, and customer conversations stay safe.
Buda runs real work on private files, internal process knowledge, and customer conversations. That only works if you trust where your data lives and what your agents can touch. Security isn't a settings page you find later — it's built into the model: every agent runs in its own isolated cloud computer, and data is scoped to the Space that owns it.
How agent execution is isolated
Each Claw executes inside its own cloud computer — a sandboxed environment with its own files, browser, terminal, and Git. Agents don't share a machine:
- Per-agent sandboxes. Every agent gets a dedicated sandbox and a named persistent volume (
buda-agent-<space>-<agent>). One agent can't read another agent's working files. - Plan-scoped resources. CPU and memory are provisioned from your Space's plan, so a runaway task is bounded by its own limits, not the whole platform.
- Idle sleep, on-demand wake. Sandboxes sleep when idle and wake when there's new work, so nothing keeps running in the background unsupervised.
Want the full mechanics — network boundaries, persistence, and cleanup? See Sandbox isolation.
Data boundaries that matter
Buda organizes everything around boundaries you control. Decide them up front:
| Boundary | What it protects |
|---|---|
| Space | A Space is a company-level container. Its Drive, agents, members, and billing are isolated from every other Space — even ones you also belong to. |
| Agent | Files and instructions scoped to one agent. Keep one agent per business function so context never bleeds across departments. |
| Channel | Which external users can reach an agent. Channels are user-scoped and route to a default Space — separate internal and customer-facing channels. |
| Session | When context continues versus resets. Sensitive conversations can start fresh instead of inheriting prior history. |
Operational best practices
- Grant the minimum. Only upload documents an agent should actually be allowed to reference.
- Separate by function. Create distinct agents for different departments, clients, or customers rather than one agent that sees everything.
- Test with non-sensitive content first. Validate a channel's behavior before pointing it at real customer data.
- Define escalation rules. Route ambiguous or sensitive cases to a human instead of letting the agent decide.
- Review uploads regularly. Drive is the explicit source of truth — keep it clean.
For regulated or enterprise use
If you handle contracts, HR records, financial data, or support transcripts, settle four questions before you scale:
- Which data is allowed into Drive
- Who can access each Space
- Which channels are approved for which audience
- What audit trail your team needs
Enterprise plans add BYOK, SSO, and self-hosted deployment for full data control. Contact us to scope it.