Custodian Labs takes you from idea to deployed, production-grade AI agent — with privacy built in at every layer. No infrastructure. No data leakage.
from custodian_labs import Custodian
model = Custodian(
model="gpt-4o",
system_prompt="You are a helpful assistant...",
)
model.deploy()Everything you need. Nothing you don't.
Guardian layer
The proprietary Guardian Layer is the guardrail that detects PII and puts you in control of how it's handled before it reaches any model.
RAG out of the box
Give your agent long-term memory and document retrieval without configuring embeddings or vector DBs. Add a knowledge base in one line and it just works.
Model-agnostic
Switch between OpenAI, Anthropic, Mistral, or local models with one line — no vendor lock-in, no rewrite required. Your agent logic stays exactly as it is; only the model underneath changes.
Zero infrastructure
No database to provision, no vector store to connect, no hosting to manage. Custodian abstracts every layer beneath your agent so you focus entirely on logic.
Research-grade foundations
Commercialised with AUT Ventures and New Zealand Government funding. Production reliability grounded in deep academic research.
No databse. No hosting. Just deploy.
Normally you'd provision a vector DB, set up a hosting environment, wire up your model provider, write retry logic, handle streaming — all before writing a single line of agent logic. Custodian collapses that entire stack to a function call.
# Step 1: Provision a vector DB + embedding pipeline
pinecone.init(api_key=...), index = pinecone.Index("my-index")
embedder = SentenceTransformer("all-MiniLM-L6-v2")
chunks = [text[i:i+512] for i in range(0, len(text), 448)]
index.upsert([(f"doc_{i}", embedder.encode(c)) for i, c in enumerate(chunks)])
# Step 2: Manual PII scrubbing (fragile regex, hope for the best)
re.sub(r"[A-Z][a-z]+ [A-Z][a-z]+", "[REDACTED]", prompt)
# Step 3: Tool dispatch + retry logic
def call_with_retry(messages, max_retries=3):
for attempt in range(max_retries):
try: return client.chat.completions.create(...)
except: time.sleep(2 ** attempt)
# Step 4: Session memory via Redis
redis_client.setex(f"ctx:{session_id}", 3600, json.dumps(messages[-20:]))
# Step 5: Flask server, auth, rate limiting, CORS, logging...
# Step 6: Write your Dockerfile, k8s config, CI/CD pipeline...
# Step 7: Finally — your agent runs. Maybe.from custodian_labs import Custodian
model = Custodian(
model="gpt-4o",
system_prompt="You are a helpful assistant...",
)
model.deploy()Assemble your team. Deploy as one.
Build teams of specialised agents that each handle their own domain — then deploy them as a single unified endpoint. Custodian's router automatically directs each query to the right agent.
Give each agent its own data sources, tools, system prompt, and topic scope. Scale the team without rewriting the interface.
the_engineer
Technology & engineering questions
the_strategist
Intelligence & strategy queries
the_scientist
Science & research topics
router
Automatically routes every query to the right agent
from custodian_labs import Agent, AgentTeam
team = AgentTeam(
agents=[
Agent(
name="the_engineer",
model="gpt-4o",
system_prompt="You are a senior engineer. Handle all tech and engineering questions.",
topics=["technology", "engineering", "AI"],
),
Agent(
name="the_strategist",
model="gpt-4o",
system_prompt="You are a strategy lead. Handle intelligence and strategy questions.",
topics=["intelligence", "strategy", "espionage", "tactics"],
),
Agent(
name="the_scientist",
model="gpt-4o",
system_prompt="You are a research scientist. Handle science and research questions.",
topics=["science", "research", "biology", "physics"],
),
],
routing_mode="single",
)
app = team.deploy()Introducing the Guardian Layer
Most privacy solutions simply strip PII — but stripping context destroys AI accuracy. The Guardian Layer detects PII intelligently, then gives you full control over what happens next, without the original PII ever leaving your envionrment.
Custom PII entities
Define your own entity types beyond the defaults — internal employee IDs, product codes, proprietary terms. Guardian Layer learns your schema.
Input (raw)
Schedule a follow-up with James Harrington at james@acmecorp.com about the contract renewal for $240,000.
Guardian Layer — Transform
Sent to AI model
Schedule a follow-up with Alex Morgan at alex@synthco.io about the contract renewal for $240,000.
Returned to you with PII relinked
Follow-up scheduled with James Harrington for Thursday at 2pm. Calendar invite sent to james@acmecorp.com.
Input (raw)
Schedule a follow-up with James Harrington at james@acmecorp.com about the contract renewal for $240,000.
Guardian Layer — Mask / redact
Sent to AI model
Schedule a follow-up with [PERSON] at [EMAIL] about the contract renewal for [CURRENCY].
Returned to you with PII relinked
Follow-up scheduled with [PERSON] for Thursday at 2pm. Calendar invite sent to [EMAIL].
Input (raw)
Schedule a follow-up with James Harrington at james@acmecorp.com about the contract renewal for $240,000.
Guardian Layer — Detect only
Sent to AI model
Schedule a follow-up with James HarringtonPERSON at james@acmecorp.comEMAIL about the contract renewal for $240,000CURRENCY.
Detection report
Person
James Harrington
99%
Email
james@acmecorp.com
100%
Currency
$240,000
97%
Org
acmecorp
88%
Transform
Replaces personally identifiable information (PII) with synthetic equivalents that preserve meaning
Input (raw)
Schedule a follow-up with James Harrington at james@acmecorp.com about the contract renewal for $240,000.
Guardian Layer — Transform
Sent to AI model
Schedule a follow-up with Alex Morgan at alex@synthco.io about the contract renewal for $240,000.
Returned to you with PII relinked
Follow-up scheduled with James Harrington for Thursday at 2pm. Calendar invite sent to james@acmecorp.com.
Mask / redact
Replaces personally identifiable information (PII) with a typed label — the fastest and most conservative transform
Input (raw)
Schedule a follow-up with James Harrington at james@acmecorp.com about the contract renewal for $240,000.
Guardian Layer — Mask / redact
Sent to AI model
Schedule a follow-up with [PERSON] at [EMAIL] about the contract renewal for [CURRENCY].
Returned to you with PII relinked
Follow-up scheduled with [PERSON] for Thursday at 2pm. Calendar invite sent to [EMAIL].
Detect only
Only surface what Replaces personally identifiable information (PII) exists, without modifying any text
Input (raw)
Schedule a follow-up with James Harrington at james@acmecorp.com about the contract renewal for $240,000.
Guardian Layer — Detect only
Sent to AI model
Schedule a follow-up with James HarringtonPERSON at james@acmecorp.comEMAIL about the contract renewal for $240,000CURRENCY.
Detection report
Person
James Harrington
99%
Email
james@acmecorp.com
100%
Currency
$240,000
97%
Org
acmecorp
88%
Simple, transparent pricing
Start today for free. Scale when you're ready.
Basic
Sandbox, try it out free
100k tokens
1k requests
GPT-4o
Email and community support
Starter
For users who want more limits, priority access, and VIP support
1m tokens
10k requests
All models + newest first
Priority 24/7 support
Direct roadmap input
Monthly 1:1 product call
Free compliance assessment
Custom
Everything tailored to your stack and compliance needs
Custom token limit
Custom request limit
On-premise or private cloud
Offline deployments
Custom compliance & data contracts
Dedicated support & SLA
From idea to deployed.
In under 10 minutes.
Get an API key, install the SDK, and have a privacy-safe AI agent running in production — no infrastructure required.