FAQ

The questions
we get most

Wiring NullRun in, how a decision is made, what happens when something fails, and how the audit log works. Plan and limit questions are on the pricing page.

  • Getting started

    How do I connect my first workflow?

    Create the workflow in /control-center/workflows, then generate an API key for it — the secret is shown once, so copy it then. From there, either wrap the function that talks to the model with @protect, or call the gate over HTTP. The quickstart is on /docs.

  • What NullRun sees

    Does NullRun see my prompts or completions?

    No. The SDK reports token counts, not text, so we never see your prompt or the completion. For a tool call it sends the tool name and that call's arguments — masked where the argument name looks like a secret — because an approval has to be bound to one specific action. Your application still calls the model provider directly. NULLRUN is not an LLM proxy.

  • Is NullRun an LLM proxy?

    No. Your application calls the provider directly. NULLRUN sits between the agent's reasoning and the side effect, not between the application and the model.

  • Policies and decisions

    Where does the policy decision happen?

    At the NULLRUN gateway, between the agent's reasoning and any consequential action. One decision point covers budget, tool permissions, rate limits, and approval flow, so a policy you did not plan for still applies.

  • What do Allow, Require approval, and Block mean?

    The gate returns one of three decisions. Allow proceeds. Require approval waits for an explicit human sign-off. Block returns a structured decision. Sensitive actions can be configured to require approval; the rest of the agent continues to run normally.

  • Can I add controls to an existing agent without rewriting it?

    Yes. Wrap the function that talks to the model with @protect — one line, and the existing graph, crew, or runner does not change. Common frameworks are recognised by the wrapper, and a custom transport gets an explicit tracking path instead.

  • What does the agent see when an action is blocked?

    A structured decision with a reason. The agent can read it, branch, and continue with the next allowed action. Blocked actions do not run.

  • What happens when a workflow hits the budget limit?

    At the cap, new actions are blocked. With a soft policy the workflow can continue inside a bounded, cumulative overdraft for the period — and only while an approval chain is active. The dashboard shows what the period spend was and how many calls were blocked.

  • Can a workflow policy weaken the organization policy?

    On caps, rate limits and blocked tools: no — organization and workflow policies merge most-restrictive-wins, and the organization's budget ceiling is enforced strictly regardless. On one thing, yes: the enforcement mode is workflow-authoritative, so a workflow-scoped Soft policy lets that workflow spend into a bounded overdraft where the organization set Hard.

  • When things go wrong

    What if my agent retries the same call after a network timeout?

    The duplicate /track is recognised as the same execution. The execution ID is minted server-side, and the settlement step is idempotent on it: a replay returns the already-recorded result instead of billing twice. Network instability becomes a retry problem. The charge is not duplicated.

  • What if the policy service is unavailable?

    The decisions that spend money or touch a sensitive action fail closed: a charge, a delete, or an external send does not proceed when the runtime cannot make a decision. The one exception is the per-key rate limiter, which is a secondary signal — if it cannot be read the call proceeds, and the budget gate still applies downstream.

  • How do I test a policy change before production?

    Replay /gate and /track against a test API key, or run them in CI with one. A key minted with the test flag is prefixed nr_test_, so it is never mistaken for a live credential. The wire contract is the same and no real LLM inference is needed — catch policy drift before it reaches production traffic.

  • Audit and proof

    How do I prove what the agent was allowed to do?

    Export the audit trail from the control center. It carries who acted, what they did, when, and the outcome — it is not a store of your conversations, so the export does not become another copy of them. The audit log is a plan feature: on Lite and Starter it is not included, so there is no trail to export.

  • Does the audit log protect against tampering?

    Every governance event is appended to a SHA-256 hash-chained sequence, and the chain can be verified on demand through the audit-log verify endpoint. The chain is sequential, not blockchain-anchored: an attacker who rewrites both a row and its predecessor's chain pointer can hide tampering. For true external tamper-resistance you would need an externally-anchored witness (HSM / public ledger) — that is not implemented. Like the export, the chain exists only on the plans that include the audit log.

  • Keys and credentials

    What happens when a key is compromised?

    Revoke the affected key. Other workflows continue running. The revoked agent stops receiving runtime decisions. For maximum safety, rotate to a new key first and drain in-flight requests with the two-phase rotation endpoint before the revoke takes effect.

  • Do I need to send my provider's API key to NullRun?

    No. The provider key stays in your application. NULLRUN never holds the provider credential.

  • Permissions across a team

    How do I restrict what my AI agent can do?

    Write a tool block list on the organization, or on a single workflow, and add approval rules for the tools that need a human sign-off. A tool call that goes through the gate is checked against that list before it runs. A disallowed tool is blocked, and the agent receives a structured block decision with the reason rather than a silent timeout. Every policy carries a version number that increments on each change.

  • How do I stop an agent from executing unauthorized actions?

    Two layers. First, the deny list: the action never reaches the tool. Second, fail-closed semantics: if the policy service is unavailable or returns no sources, the runtime refuses the action instead of assuming it is permitted. An agent cannot fall through to a side effect when authorization cannot be established.

  • How do I manage AI agent permissions across a team?

    Policies are written once at the organization level and inherited by every workflow. A workflow policy can add restrictions, and on caps, rate limits and blocked tools the combination resolves most-restrictive-wins. The one dimension a workflow can loosen is the enforcement mode, which is workflow-authoritative. Manage access control from /control-center/policies.

Didn't find your answer?

Send us a note and we'll get back within a working day.
For security reports, use support@nullrun.io instead.

    FAQ — NullRun