AI agents can delete the database, spam your entire CRM, leak your API keys, take down the site, rack up cloud bills, share private files, violate compliance, blow your API budget, email every customer, spend way too much, deploy broken code.Trust is the mistake.
Your agent ran an action nobody approved, against a payload nobody read.
NullRun authorizes the action first and binds that approval to the payload the agent is about to send, not to the prompt that asked for it.
Works with the stack you already run
- OpenAI
- Anthropic
- Google Gemini
- Mistral
- Cohere
- LangChain
- CrewAI
- LangGraph
- AutoGen
Compatibility examples. Marks belong to their respective owners.
Control the actions.
Control the cost.
Keep the evidence you acted on.
From policy enforcement to a full audit trail and decision history.
All of it runs in the control center.
Prevent unnecessary spend before it happens.
Require approval when actions carry risk.
Know exactly why an action ran.
Decorate a function.
Operators govern it.
Operators author the policy in the dashboard. Devs can't accidentally change the rule.
Make every decision
in the control plane
Straight from the control center: dashboards, audit views, and the policy editor.
A real request and response. Field names and units are the wire contract (/api/v1/gate).
- allow
- require_approval
- block
POST /api/v1/gate- X-NULLRUN-PROTOCOL
- 4
- tool
- stripe.refund
- business_impact
- { "kind": "financial", "amount_cents": 12000, "currency": "USD" }
200 OK- decision
- require_approval
- execution_id
- 0193f2c4-7a1e-7b3d-9c05-2f6b8ad41e77
- action_digest
- dfc96387ca539b71…f0df27
- policy_version
- 7
- projected_cost_cents
- 310
- remaining_budget_cents
- 4090
execution_id is minted by the server, not the SDK. action_digest binds the grant to this exact payload, so it cannot be spent on a different amount.
When the agent goes off-script,
what stops it?
Same agent, same workflow. What changes is whether anything stands between it and the outside world.
- Runaway loop
Spend explodes
The same tool, retried every iteration, until the card declines.
Loop blocked
Repeated calls to one tool trip the circuit breaker; the next call on that workflow is refused.
- Sensitive action
Executes anyway
Refunds, deletes, customer emails. All of it runs immediately.
Human approval required
An operator decides, and that decision is recorded.
- Credentials
Reaches logs
Tool arguments land in plaintext, unredacted.
Redacted at the gate
Credential-shaped strings are stripped before the row is written.
- Decision history
No reliable history
Weeks later, no record survives the retention window.
Blocks and approvals recorded
Every block and approval is appended to a SHA-256 chain.