NullRun is the runtime decision layer for tool-using AI agents. Drop in a one-line decorator, set a budget, and govern cost, tool use, and workflow state before a runaway agent turns into an incident.

Managed runtime control plane. Not a self-hosted deployment.

• Stop runaway costs. Hard budget policies stop a workflow at the cap. Soft policies can allow a bounded overdraft for an active chain — but only when the policy uses enforcement_mode = Soft, an active chain_id exists, and the projected cost stays within the configured overdraft limit.
• Action-bound approvals. Approval rules can match against a typed action payload (a money amount or a tool-call argument bag). Every approval is bound by a SHA-256 digest of the exact payload — the gate refuses to honour the grant if a later /execute arrives with a different amount or different arguments.
• Control from the dashboard. Pause or kill a workflow through the WebSocket control plane. Signals are applied at the next gate or yield boundary.
• Govern sensitive tools safely. Database writes, shell, file deletes, and external actions can be blocked by ToolBlock policies or routed through human approval, with a full audit trail.

The Python SDK is the only runtime client for the agent hot path. Integration coverage varies by framework:

• End-to-end tested (decorator plus dedicated behaviour tests): LangGraph, CrewAI, AutoGen, LlamaIndex.
• HTTP transport tested: OpenAI.
• Decorator implemented, no dedicated end-to-end test: OpenAI Agents, LangChain, Anthropic.
• Extractor unit-tested, no full transport-to-track integration test: Mistral, Gemini, Cohere, AWS Bedrock.

• nullrun.io — managed control plane and dashboard in one place. Start with the Lite plan, no credit card.
• pip install nullrun — Python SDK (0.14.x). Pass your nr_live_... API key to init() to start tracking and enforcing policy.
• docs.nullrun.io — install, quickstart, concepts, and recipes.

NullRun evaluates structured action requests before execution and returns allow, block, or require_approval. It does not inspect prompts, tool arguments, or model output semantics. Cost enforcement relies on SDK-reported estimates and usage — a malicious SDK that controls its own cost reports is not protected by the gate.

• nullrun-sdk-python — the Python SDK (pip install nullrun).
• nullrun-docs — the documentation site.
• nullrun-examples — copy-paste-runnable examples for the most common agent frameworks.

• Bug or feature? Open an issue in the relevant repo.
• Security? See SECURITY.md — please don't file public issues.
• General questions: Discord is the fastest channel.
• Email support: support@nullrun.io.