Custodian Labs
A Python toolkit for deploying AI agents in a handful of lines of code that automatically detects and masks personal data (like names, emails, or ID numbers) before it reaches a model.
🔗 Visit Custodian Labs
Description
Building an AI agent that handles real customer data creates an uncomfortable question most teams solve poorly: what stops a customer's email address or ID number from ending up verbatim in a prompt sent to an external LLM provider? Custodian Labs bakes an answer directly into its SDK — a "Guardian Layer" that automatically detects personal data in text and either masks it, transforms it, or just flags it, depending on how you configure it, before the agent's logic ever touches it.
Beyond privacy, it's designed to get a working AI agent running in about five lines of code, with built-in retrieval-augmented generation that doesn't require separately standing up and configuring a vector database, multi-agent routing for handing off tasks between specialized agents, and support for OpenAI, Anthropic, Mistral, or local models interchangeably.
💬 Our review
The short version: Custodian Labs is a genuinely early-stage but credible attempt to make privacy-by-default a built-in feature of agent deployment rather than something bolted on afterward, which most general-purpose frameworks don't offer out of the box.
Frameworks like LangChain give you far more flexibility and a much larger ecosystem, but they don't include automatic PII detection and masking as a first-class feature — with Custodian Labs, you get that plus RAG-without-vector-DB-setup in exchange for a smaller, newer, less battle-tested toolkit. It's backed by pre-seed funding from AUT Ventures and New Zealand government support, which is a real signal of legitimacy, but a small GitHub star count and sparse public documentation mean it's genuinely early — worth prototyping with before betting a production system on it.
💰 Pricing
📊 Global score
🤖 AI-enriched data
Gratuit : 100k tokens / 1 000 requêtes. Starter : 19 $/mois (1M tokens / 10 000 requêtes + support prioritaire). Enterprise : sur devis, détails non publiés.
Pros
Déploiement d'un agent en environ 5 lignes de code
Détection et masquage automatique des données personnelles intégrés nativement
RAG intégré sans configuration séparée de base vectorielle
Financé en pre-seed par AUT Ventures et le gouvernement néo-zélandais
Cons
Stade très précoce (pre-seed), peu d'historique public
Faible visibilité GitHub (14 étoiles), signal de jeunesse plutôt que d'adoption
Documentation publique limitée (pas de /docs ou /blog visibles)