Now in Early Access · Built for Europe

Your agent changed.You didn't know.

Driftbase fingerprints how your AI agent behaves in production. One integration — instant alerts the moment behavior silently shifts. Privacy-first by design.

driftbase — production diff · v1.0 → v2.0
# 1. Install the open-source SDK$ pip install driftbase# 2. Wrap your agent — one linefrom driftbase import track@track(version="v1.0")# 3. Diff after 50+ runs$ driftbase diff v1.0 v2.0⚠ BEHAVIORAL DRIFT DETECTED drift_score 0.73 ← above 0.20 threshold decision_drift 0.81 ← escalation rate +3.2× latency_drift 0.31 ← p95 +180ms error_drift 0.04 ← stable Computed in 120ms · No data left your machine$

The problem

Flying blind after every deploy.

Your observability stack tells you an agent ran. It doesn't tell you it behaved differently than last week. A model provider silently rolls an update — error rate stays flat, agent breaks anyway. You find out three days later, from a customer.

01

No behavioral signal

Latency and error rate monitoring misses how your agent makes decisions.

02

Days to detect regressions

Teams find behavioral regressions from customer complaints — not alerts.

03

Tests can't catch model updates

Unit tests check your code. They can't catch a provider-side model change that rewired how your agent reasons.

04

No audit trail for regulators

EU AI Act and DORA require behavioral records. Most teams have none and can't reconstruct them after the fact.

How it works

Running in
20 minutes.

No test cases. No golden datasets. Driftbase learns "normal" from live production traffic — then tells you when it changes.

01 — CAPTURE

One line of code

Callback handler or HTTP proxy. LangChain, LangGraph, AutoGen, CrewAI, or raw API calls — all supported.

handler = DriftbaseCallbackHandler(
version="v1.0"
)
02 — FINGERPRINT

Behavioral baseline

After 50 runs, a fingerprint is computed: tool distribution, latency percentiles, decision rates, error patterns. Zero raw content stored.

tool_dist → {search: 78%}
p95_latency → 423ms
escalation_rt → 12%
03 — DIFF & ALERT

Deploy with confidence

Every deploy triggers a fingerprint diff. Drift above threshold fires Slack, blocks the GitHub PR, and writes a timestamped audit record.

SIGNIFICANT score: 0.73
Slack · PR blocked · audit log

Features

Everything your team needs.
Nothing it doesn't.

Behavioral monitoring

Four dimensions of drift

Decision rate, tool distribution, latency profile, error pattern — each tracked independently so you know exactly what shifted.

PRODUCTION · v2.0 vs v1.0LIVE
decision_drift
0.81HIGH
latency_drift
0.31MED
tool_dist
0.17LOW
error_drift
0.04STABLE
AI-powered analysis

Root cause in plain English

Don't just know that drift happened — know why. Driftbase analyzes fingerprint deltas and explains what changed and where to look first.

"Your agent escalates to human support 3.2× more often since v2.0. The search_knowledge_base tool returns results but the agent doesn't resolve with them — the reasoning step changed, not the tool. Investigate system prompt resolution instructions."
Compliance · Built for Europe

GDPR by architecture.
Not by policy.

Built for European developers who need behavioral intelligence without compromising privacy. Content is hashed on capture — structurally incapable of leaking personal data. No raw inputs or outputs ever leave your infrastructure.

Designed to satisfy EU AI Act Articles 9 & 12, DORA change documentation, and GDPR by default. Your DPO will thank you.

Compliance coverage

Five frameworks.
One integration.

Zero raw content storage — hashed on captureGDPR
Art. 9 & 12 behavioral audit exportsEU AI Act
Behavioral change documentation, time-stampedDORA
Full VPC self-hosting — data never leaves your infraOn-prem
Regulator-ready exports on demandAudit
Temporal monitoring

Catch silent provider updates

Your code didn't change — but OpenAI or Azure silently updated the model. Driftbase's rolling baseline detects the shift within hours.

POST /diff/temporal
{ baseline_hours: 168, current: 12 }
⚠ TEMPORAL DRIFT 0.42
→ provider model update · 14:32 UTC
Privacy architecture

Hashes, not conversations

Users' messages never enter our pipeline. Behavioral signals are extracted and hashed before leaving your environment.

# Never stored
"Refund order #92847-B"
# Captured instead
sha256(input) → a3f8c21d
tool_calls["lookup","refund"]
latency_ms387

Before vs after

Your team, two ways.

Without Driftbase
With Driftbase
Agent behavior changes silently after every model update
Behavioral fingerprint computed automatically from live traffic
You find out from a customer complaint, 3 days later
Slack alert and PR block within 2 hours of any behavioral shift
Manually trace logs to guess what changed and when
Plain-English root cause — exactly what changed and why
Every production deploy is a roll of the dice
Diff score tells you it's safe to ship before it ships
No audit trail — regulator asks, you have nothing
Time-stamped behavioral records generated automatically
US tools store your users' raw conversations by default
Zero raw content stored — GDPR compliant by architecture

Pricing

Free to start.
No catch.

The core SDK is open source and completely free. Install it, run it on your real agent, and see what behavioral monitoring actually looks like — before deciding anything else.

Team plans with cloud dashboard, Slack alerts, and EU compliance exports are coming soon.

Get notified when Pro launches →
Free forever
€0 always

No account. No credit card. No artificial limits.

  • Full open-source SDK — Apache 2.0 licensed
  • Local SQLite fingerprinting — data stays local
  • CLI drift reports and deployment diffs
  • LangChain, LangGraph, AutoGen, CrewAI support
  • HTTP proxy for framework-agnostic capture
  • Up to 1,000 monitored runs / month
Get Started

Questions? Reach the team

Stop flying blind
after every deploy.

Start monitoring behavioral drift today. Get your baseline in 50 runs. Know the moment something shifts — before your users do.

Open-source · GDPR by architecture · Built for European developers