Everything you need to build enterprise AI agents with cryptographic identity, capability governance, HITL controls, and immutable audit trails.
From install to your first governed agent in 5 minutes. Install the SDK, register an actor, run it, and verify the audit trail.
Start hereActor, Model, Tool, Workflow, and HITL — the five composable building blocks. Start here if you're evaluating the SDK architecture.
Explore primitivesThe five ideas that underpin everything: agent identity, capability explicitness, immutable observation, governed workflows, and tenant isolation.
Learn the conceptsStep-by-step tutorials for common patterns: add HITL to a workflow, migrate from LangChain, integrate with Azure OpenAI, compliance setups.
Browse guidesTypeScript SDK, Python SDK, and REST API — full auto-generated reference with types, examples, and response schemas.
View API referenceCulvii Cloud, self-hosted Docker, BYOC (your VPC), and air-gapped. Configuration reference for every deployment model.
Deployment guidePrerequisites: Node.js 18+ or Python 3.11+. A Culvii account — start your 30-day trial.
npm install @culvii/kit
npx culvii init # Creates .culvii/ and provisions your API key
import { Actor, Model } from '@culvii/kit' const agent = new Actor({ name: 'my-first-agent', model: Model.openai('gpt-4o'), capabilities: ['read:documents'], }) console.log(agent.did) // → did:culvii:my-tenant:my-first-agent-a3f9b2...
const result = await agent.run({ prompt: 'Summarize the Q1 financial report.', context: { document: '...' }, }) // Every step is signed, hashed, and logged
import { Auditor } from '@culvii/kit' const proof = await Auditor.verify({ agentDid: agent.did }) // → { valid: true, events: 3, // merkleRoot: 'a3f9b2...', tampered: false }
Each primitive maps to a real engineering concern — not an academic abstraction. They compose.
Identity-bearing execution unit. Wraps model + tools + capabilities.
Provider-agnostic LLM configuration. OpenAI, Anthropic, Azure, or local.
Typed, capability-scoped function. All calls logged and signed.
Governed execution graph. Sequential, parallel, conditional.
Human-in-the-loop checkpoint. Suspends, waits for approval, resumes.
Agents aren't just function calls or prompt chains. They're named, identity-bearing entities registered in your tenant — accountable in the same way a service account is accountable.
Agents can only do what they've been explicitly granted. No capability, no execution — regardless of what the LLM is capable of. Capabilities expire, rotate, and revoke instantly.
You don't need to add logging. Every tool call, model invocation, and workflow step emits an observation event that flows through a hash chain anchored to a Merkle Tree root.
A Culvii Workflow is a governed execution context with checkpoints, HITL gates, and rollback semantics. Pause execution, require approval, and resume — without losing state.
Everything — agents, capabilities, keys, logs — is scoped to a tenant. Tenants never share data, keys, or audit logs. Full isolation at the database, application, and cryptographic layers.