NILScript
OpenAPI for agent actions, so hallucinations can't write
Last verified:
What is NILScript?
NILScript is an open standard for how AI agents act on real business systems safely. It provides Network Intent Layer (NIL) plus the nilscript DSL—a neutral wire contract between the agent that decides and the system that owns the truth. Every write follows propose → approve → commit → rollback, ensuring nothing touches your data until a human explicitly approves, and agents can only name verbs your backend actually declares.
Key features include structural defense against hallucinations and prompt-injection attacks, with 0.00% unauthorized writes across 4,216 real evals on two models. The CLI toolkit lets you build and verify adapters with three core files: system.py for I/O, translate.py for verb-to-native translation, and compensation.py for reversibility. The Playground shows real propose→commit traces in seconds.
NILScript is designed for developers building AI agents that interact with production systems, engineering teams integrating agents with backends, and organizations needing safe agent automation. It works in any language against any backend, serving as the 'USB for software'—the neutral standard for connecting systems to agents.
NILScript pricing
Pricing model: Freemium
Free and open source. No account, API key, or waitlist required. The CLI is completely free with no paid tiers. Install, scaffold, and use the Playground without any cost.
NILScript pros
- 0.00% unauthorized writes across 4,216 real injection evals
- Structural defense not dependent on specific AI model
- Every write requires human approval before committing
- Agents can only use verbs your backend explicitly declares
- Hallucinations cannot write to your data
- Works in any programming language
- Supports any backend system
- No account, API key, or waitlist required
- Free and open source with no cost
- CLI toolkit for building and verifying adapters
- Three-file surface you build once, any NIL-speaking agent works
- 100% benign tasks still complete successfully
- Real trace visualization in Playground
- Closed wire contract prevents unauthorized access
- Prompt-injection attacks cannot hijack writes through NIL
NILScript cons
- Requires backend to explicitly declare available verbs
- New standard still at version 0.3.0 (not final)
- May require adapter development for existing systems
- Only supports write operations through propose-approve-commit-rollback
- Limited to systems that can implement NIL adapters
- No built-in model selection or inference
- Requires developers to scaffold adapters manually
- Documentation may be incomplete for advanced use cases
Frequently asked questions about NILScript
What is NILScript?
NILScript is the open standard for how AI agents operate real business systems safely, in any language, against any backend. It consists of two layers: NIL (Network Intent Layer)—a closed wire contract—and the nilscript DSL. It acts as the 'USB for software,' providing one neutral wire contract between the agent that decides and the system that owns the truth.
How does NILScript prevent unauthorized writes?
Every write goes through propose → approve → commit → rollback. Nothing touches your data until a human explicitly says so. Agents can only name verbs your backend actually declares, making the defense structural rather than model-dependent. This achieved 0.00% unauthorized writes across 4,216 real prompt-injection attacks.
What is the benchmark performance?
Across 4,216 real injection evals on two models, raw agents were hijacked into a write on up to 1 in 22 cases (4.46% hijack rate). Routed through NIL, unauthorized writes commit at 0.00%, while 100% of benign tasks still complete successfully.
How do I get started with NILScript?
You can get up and running in three commands using the CLI. Install the CLI, explore and scaffold the three adapter files (system.py, translate.py, compensation.py), and watch a real propose→commit loop in the Playground. No account, key, or waitlist is required.
What are the three core adapter files?
The three files that become yours are: system.py (the one place where I/O happens), translate.py (handles verb to native translation), and compensation.py (manages reversibility for rollback operations). Everything else is generated and identical across adapters.
Does NILScript work with any backend?
Yes, NILScript works against any backend and in any programming language. It serves as the layer between the agent and the backend, providing a neutral contract that any NIL-speaking agent can use against your system once you build the adapter surface once.
What version is NILScript currently at?
NILScript is currently at version v0.3.0, indicating it is an open standard that is still in development and not yet at final release.
Can agents bypass human approval with NILScript?
No. Nothing touches your data until a human explicitly says so. The propose → approve → commit → rollback flow ensures human approval is mandatory before any write commits, making it impossible for agents to bypass approval even during prompt-injection attacks.
Is NILScript free to use?
Yes, NILScript is completely free and open source. There is no account required, no API key needed, and no waitlist to join. You can install the CLI, scaffold adapters, and use the Playground without any cost or payment.
What happens if an agent hallucinates with NILScript?
Hallucinations cannot write. Since agents can only name verbs your backend actually declares and every write requires human approval through the propose → approve → commit → rollback flow, hallucinated commands simply cannot touch your data.