OneCLI

Product Introduction

1. Definition: OneCLI is an open-source, self-hosted agent harness and gateway platform for deploying secure, sandboxed AI assistants within an organization. It functions as a policy-controlled orchestration layer that sits between AI agents (like those powered by LLMs) and a company's internal tools and data.
2. Core Value Proposition: OneCLI exists to enable safe, scalable, and secure AI agent adoption for teams. It solves the critical security and compliance challenges of giving AI models direct access to production systems by enforcing hard policy limits outside the AI's control, ensuring agents never hold real credentials and operate strictly within predefined, human-approved boundaries.

Main Features

1. Secure, Credential-less Agent Gateway: The core technical architecture is a centralized gateway through which all AI agent requests must pass. It operates on a just-in-time, scoped credential injection model. When an agent needs to perform an action (e.g., query a database), the request is routed through the OneCLI gateway, which dynamically injects the necessary, least-privilege credentials for that single operation. The agent itself never stores or has persistent access to any real secrets, drastically reducing the attack surface and credential leakage risk.
2. External Policy Enforcement Engine: Unlike mere prompt instructions, OneCLI implements enforceable guardrails as code outside the AI model's runtime. This system acts as a hard firewall, intercepting and evaluating every agent action against a centralized policy rulebook. Rules can define absolute prohibitions (e.g., "never delete a repository"), rate limits (e.g., "slow down after 50 API calls/minute"), and mandatory human-in-the-loop approvals for sensitive operations (e.g., "pause and ask before emailing a customer").
3. Per-Employee, Sandboxed Agent Instances: OneCLI provisions a dedicated, isolated agent environment for each user. This sandboxing ensures strict access control inheritance; an agent can only interact with the tools, data, and accounts that its assigned human user already has permission to access. This prevents lateral movement, ensuring a support agent cannot accidentally or maliciously access billing or payroll systems.

Problems Solved

1. Pain Point: The high risk and lack of control in deploying autonomous AI agents that can directly execute commands in production environments. Traditional methods rely on the AI's compliance with prompt-based rules, which can be ignored or bypassed ("prompt injection"), leading to catastrophic actions like data deletion or unauthorized transactions.
2. Target Audience: DevOps and Platform Engineering teams tasked with AI infrastructure security; Security and Compliance officers (CISOs, GRC teams) in tech companies; Engineering managers and team leads looking to safely boost productivity with AI automation.
3. Use Cases: Automating routine engineering tasks (creating tickets, checking logs, deploying to staging) safely. Allowing customer support agents to safely query internal systems to resolve tickets without credential sharing. Enabling marketing or sales teams to automate data enrichment and CRM updates while maintaining strict data governance and approval workflows.

Unique Advantages

1. Differentiation: Unlike AI chatbot platforms or RAG systems that only retrieve information, OneCLI is built for safe execution. Compared to building in-house agent security, it provides a pre-built, open-source platform. Versus using agents via direct API keys, it eliminates credential persistence in the agent layer.
2. Key Innovation: The decoupling of policy enforcement from the AI model itself. By moving the "rules engine" into a separate, privileged gateway layer that sits in the execution path, OneCLI guarantees that policy violations are physically blocked before an action reaches a target system, making safety a matter of system architecture rather than model behavior.

Frequently Asked Questions (FAQ)

1. Is OneCLI an AI model? No, OneCLI is not an AI model. It is a security and orchestration platform (a "harness") that allows you to safely run and manage existing AI agents and models (like GPT-4, Claude, or open-source LLMs) by controlling their access to tools and enforcing security policies.
2. How does OneCLI handle authentication and credentials? OneCLI uses a dynamic credential injection model. Real credentials are stored in a secure vault. When an authenticated agent makes an authorized request, the OneCLI gateway temporarily injects the minimal necessary credentials for that specific action. The agent itself never sees or stores the permanent secret.
3. Can OneCLI be self-hosted on private infrastructure? Yes, OneCLI is designed as a self-hosted, open-source platform. This allows organizations to deploy it within their own private cloud or on-premises environment, ensuring all data, agent activities, and the policy engine remain inside their security perimeter.
4. What happens if an AI agent is instructed to break a rule? The external policy engine in the OneCLI gateway will intercept the attempted action and block it. For example, if a rule forbids deleting repositories and an agent is prompted to do so, the gateway will stop the request before it reaches the version control system's API, regardless of what the AI model decides to do.
5. What integrations does OneCLI support? OneCLI provides a gateway that can integrate with any tool that has an API (like GitHub, Slack, Jira, Datadog, CRM systems). Its open-source nature allows teams to build custom connectors. It delivers these integrated agents through user-facing interfaces like Slack and a web application.