Frequently Asked Questions
How does Unbound integrate with existing AI tools?
Your developers keep using their own tools, providers and subscriptions. Unbound governs what those tools do: terminal commands, file writes and MCP tool calls are allowed, warned on or denied before they run. Governance covers Claude Code, Codex, Cursor, Copilot, Augment Code and Pi (Beta), and Cowork, which reports under its own name. Discovery inventories more — around two dozen tool families across macOS, Windows and Linux — so you can see what is installed even where policy cannot yet be enforced. Roo Code, Cline, Kilo Code and Custom Access are set up differently: point them at Unbound with the keyunbound tools connect <TOOL_TYPE> prints.
How does Unbound handle sensitive information?
Data policies inspect prompts for personal data and secrets, and each guardrail is set to Block — the request is stopped before its content reaches a model — or Audit, which records the detection and lets the request through. Every hit lands in Analytics → Data, with the tool, the user, the guardrail and the action. Prompt and response content is recorded so Logs can show a session end to end, and nothing in the product deletes it as it ages. Data Retention sets out what is kept, what expires on its own, and what you can delete.What does Unbound detect in a prompt?
- Secrets — API keys and tokens, database connection strings and cryptographic keys. All three categories are covered unless you narrow it.
- Personal data — Credit Card Number and Email Address, with more identifier types opening up for policy configuration over time. Where your organization has PII policies switched on, the create-policy form shows the current set.
What happens if Unbound is unreachable?
You choose. Settings → Policy Enforcement sets what your agents do when the policy check cannot run: Allow, so work continues uninterrupted — the default — or Block, so work stops until Unbound is reachable again.Can I bring my own models and provider keys?
Yes, for the traffic Unbound routes on your behalf. Your developers keep their own provider accounts and subscriptions; Unbound governs the tools rather than selecting models. Custom and fine-tuned deployments are addressable where your provider exposes them — Vertex AI and Bedrock get their own per-model entries. For agents like Claude Code and Cursor, model choice stays in the tool and on your own subscription. Unbound observes and governs it rather than selecting it.Can I host Unbound on my own?
Unbound runs as a managed service. Enterprise tenants can be deployed on your own hostnames, and the CLI points at them withunbound login --domain your.domain or unbound config urls. Talk to us about what your environment needs.
What can Unbound tell me about cost?
Estimated spend is attributed per user, per model and per tool, broken down by department and manager, over any window up to 90 days.- Budgets and spend limits cap what one person can spend, monthly or over a rolling window, and block once crossed.
- Recommendations flag where a different Claude Code subscription plan would cost your organization less, with the saving calculated from actual usage.
What analytics and insights does Unbound provide?
- Spend, tokens and agent activity across the organization, filtered by department or manager
- Per-developer activity — tasks, cost, tokens, models, and how much of the work ran unattended
- Active Users, Tokens Used, Agentic Users and Peak User (Tokens), so you can see both the spread and the outliers
- File-level analytics: which files receive the most agent-written code
- Adoption over time — weekly and monthly active users, an adoption funnel, and take-up per team
- Recommendations covering risky commands, policy gaps, stale devices and tampered setups
Is there a free trial available?
Yes. New organizations start on a 14-day trial. Free credits at signup apply to a specific set of free-tier models. When the trial ends, requests are refused with “Your trial has ended” until a subscription is in place.How does Unbound ensure data privacy?
- A blocked action is stopped before it runs
- Personal data and secrets detection, set to block or audit per guardrail
- A generated export carries prompt text, and is deleted six hours after it is built
- Role-based access: built-in roles from Admin to Member, plus custom roles that set View or Edit per area - Analytics, Logs, Policies, Configuration - and whose data they cover