envariant
Now in private beta · 2026

The control layer for foundation models

Envariant is an interpretability SDK that lets your team inspect, steer, and control model behavior — from prototype to production, in a few lines of code.

Access is whitelist-only while we onboard teams in batches.

Trusted by AI teams shipping models to production

Northwind AI
Atlas Labs
Verdant
Cobalt
Meridian
Substrate

The product

Understand what your model is doing — and decide what it does next

Envariant turns opaque foundation models into systems you can reason about, debug, and govern with the same rigor as the rest of your stack.

Inspect

Surface the internal features driving a model's output. Attribute decisions to concepts, neurons, and prompts you can actually read.

Steer

Nudge behavior at inference time with activation steering. Dial concepts up or down without retraining or brittle prompt hacks.

Control

Set guardrails as code. Define invariants the model must hold and enforce them on every request in production.

Live tracing

Stream feature activations per token in real time, with low-overhead hooks built for production traffic.

Safety invariants

Catch regressions before users do. Assert behavioral properties and get alerted the moment a model drifts.

Model-agnostic

One SDK across open weights and hosted APIs. Swap providers without rewriting your interpretability layer.

How it works

From black box to control surface in three steps

01

Instrument

Wrap any model call with the Envariant SDK. Activations and attributions start flowing — no infra changes required.

02

Interpret

Explore the feature dictionary in the dashboard. See which concepts fire, when, and why for any input.

03

Govern

Write steering rules and safety invariants as code, then ship them to production with confidence.

The SDK

Drops into your codebase. Speaks your language.

A typed, ergonomic SDK that adds an interpretability and control layer to any model call — without rebuilding your inference pipeline.

  • First-class TypeScript and Python clients
  • Low-overhead hooks designed for production traffic
  • Invariants and steering rules versioned as code
control.ts
import { Envariant } from '@envariant/sdk'
 
const ev = new Envariant({ model: 'gpt-frontier' })
 
// inspect which concepts drove the output
const { features } = await ev.trace(prompt)
 
// steer behavior at inference time
ev.steer('verbosity', -0.4)
 
// enforce an invariant in production
ev.assert('no_pii_leakage')

Pricing

Simple plans. Invite-only for now.

Purchasing unlocks once your team is approved off the waitlist. Pick a plan and request access — we'll route you to checkout.

Developer

$0/ month

For solo builders exploring interpretability.

  • 1 project
  • 10K traced requests / mo
  • Inspect + steer
  • Community support

Team

Popular
$499/ month

For teams shipping models to production.

  • Unlimited projects
  • 5M traced requests / mo
  • Production invariants
  • Live tracing & alerts
  • Priority support

Enterprise

Custom

For orgs with scale and compliance needs.

  • Volume tracing
  • Self-hosted deployment
  • SSO & audit logs
  • SLA & dedicated support

Take control of your foundation models

Join the waitlist to get early access to the Envariant SDK and dashboard. We're onboarding teams in batches.