OneCli
onecli is an open-source credential vault for AI agents. It acts as a runtime broker between agents and external services: credentials are registered with onecli once, and at runtime onecli injects secrets into tool calls without exposing keys in the agent's conversation context, prompt, or config files.
AI agents that call external APIs — pushing commits, sending messages, querying databases — need credentials to do so. The common approach is to place keys in environment variables, config files, or directly in prompts. All of these make the key visible in the LLM's context window and potentially in conversation logs. If the agent is compromised via prompt injection or if transcripts are stored, the key is exposed.
onecli separates credential storage from agent execution. The agent requests access to a named service; onecli authenticates the request and injects the secret at the point of use without it appearing in the conversation. This mirrors the pattern of secrets managers in CI/CD pipelines (HashiCorp Vault, AWS Secrets Manager) but is designed specifically for the agent invocation model.
Placed in Assess at inner position because credential leakage through agent context windows is a real and underappreciated risk surface. As agents gain access to more external services, a dedicated credential broker becomes an important part of the agent security stack. The inner position reflects that teams building agents with external service integrations should be evaluating this pattern now, even if onecli itself is early-stage. The project is in the same security design space as the constrained tool environment pattern embodied by execute_code_py.
