Custodian Labs AI Agent

Custodian Labs provides a Python-first platform for building and deploying AI agents without managing the underlying infrastructure. The core workflow, as described on the vendor page, is four lines: import the Custodian class, pass a model name and system prompt, call deploy(). The platform handles hosting, vector storage, embedding pipelines, retry logic, and model routing underneath — none of it is exposed unless you need it to be.

The differentiating feature the vendor calls out explicitly is the Guardian Layer — a proprietary PII detection and handling system that intercepts inputs before they reach any model call. Unlike regex-based scrubbing, which the vendor’s own comparison code labels ‘fragile, hope for the best,’ the Guardian Layer is described as a structured control point where developers decide how detected PII is handled. For teams building agents that touch customer data, this removes a compliance step that would otherwise require a separate pipeline.

Custodian fits teams that want agent logic deployed quickly without a DevOps specialist — Python developers who know what they want the agent to do but don’t want to maintain the stack beneath it. It also fits multi-provider workflows: the docs describe switching between OpenAI, Anthropic, Mistral, or local models with a one-line change, leaving agent logic untouched. Where it breaks is any scenario requiring infrastructure visibility — custom chunking strategies, self-hosted deployment for data residency requirements, or audit-level access to what the vector store is doing. The platform is cloud-hosted only; no self-host binaries or containers are described anywhere in the vendor documentation.

An API key is available to get started. The platform is not open-source, though the vendor links to a GitHub presence. Backing from AUT Ventures and the New Zealand Government is cited as grounding for production reliability claims.