Konstera
The platform

One layer. Every capability an agent needs to be useful.

Ontology, context graph, multi-LLM router, kernel services, human-in-the-loop, policies, metering, audit — under one roof, with multi-tenancy and durable execution from the first line.

The three planes

Where Konstera lives in your stack.

Plane
Experience

Portals, chat surfaces, embedded UIs, mobile — wherever an agent's output needs to land.

Plane
Operational / Agent
Konstera

Konstera. The nervous system between what you know and what you do. The missing plane.

Plane
Integration

iPaaS, master data, API gateways — the pipes we read from and write to.

Plane
Data / Analytical

Warehouses, lakehouses, vector DBs — the snapshots and embeddings we consume.

Plane
Systems of Record

Shopify, NetSuite, Zendesk, Klaviyo, M365 — the authoritative state of the business.

The analytical plane was Snowflake's land-grab. The transactional plane was built decades ago. The operational / agent plane was a mess of iPaaS + custom code + RPA — until agents made it a coherent product category. That's where Konstera lives.

Kernel services

What the kernel gives every agent, for free.

Nine services exposed to agents via Model Context Protocol. The agent speaks MCP; the kernel handles tenancy, metering, auth, auditing underneath.

Context graph

Shared live view of customers, orders, products, inventory. Identity resolution across connected systems with per-field provenance.

Ontology

Canonical entities (Customer, Order, Product, Return…) with explicit shadow mappings into every connector. One schema, many sources.

Multi-LLM router

Claude · OpenAI · Gemini · DeepSeek behind one call. Prompt caching, fallback chains, per-model pricing frozen at call time.

Durable execution

Every invocation is a Temporal workflow. Retries, timeouts, cron, sagas, webhooks, HITL signals — first-class, not hand-rolled.

Human-in-the-loop

Thread-first reviewer inbox. Four actions per message. Autonomy as scope-stacked policy, not a global toggle.

Cost governance

Token-level attribution per tenant · agent · user · model. Budgets enforced before each LLM call. 7/14/30-day trends in-portal.

State & memory

Per-agent and per-tenant state stores. Short-term working memory and long-term persistent memory via one API.

Credentials

Scoped, auditable, encrypted at rest. Agents never see raw secrets — the kernel brokers access per invocation.

Vector & RAG

pgvector namespaces per agent, permission-aware retrieval. Knowledge is a connector type — bind docs to an agent, not a tenant.

Human-in-the-loop

Humans and agents, as peers.

Most AI platforms treat approvals as a bolted-on queue — reviewers get a worse UX than their email client. Konstera treats human review as a first-class product. Agents draft; reviewers decide; every action feeds a policy that tunes autonomy over time.

  • Thread-first UX
    Email, support ticket, chat session, agent output — all become a thread, in one surface, with full history.
  • Four actions per message
    Send as-is · Edit & send · Ask AI to redraft · Reject with reason. Logged — feeds the reject-rate feedback loop.
  • Autonomy as scoped policy
    Scope-stacked rules: binding › agent › tenant › platform. Threshold-gated. Change behaviour without redeploying.
  • Replayable, auditable
    Every invocation is durable. Replay the exact state the agent saw when it drafted the message.
Autonomy policy — live
evaluating
Agent draft
“Where's my order?”
confidence 0.96standard
platform: require_review for all messagesreview
tenant: allow_autonomous when confidence > 0.9auto
agent: customer-service — require_review for refunds > £50no match
binding: high-value — always require_reviewno match
Narrowest matching scope wins. Auto-sent · tenant
Governance & security

Enterprise controls, built in from the first line.

Multi-tenancy

PostgreSQL row-level security on every tenant-scoped table. Application role has NOBYPASSRLS. Verified in schema, not promised in prose.

SSO · OIDC + Entra

Per-tenant IdP. Group-claim → role mapping. JIT provisioning. Native fallback for dev and break-glass.

Observability

Structured logs + OpenTelemetry. Workflow replay for every invocation. Portal log viewer with trace drill-down.

Delivery parity

Same images in SaaS, Helm, Compose. No feature-gated enterprise tier. Self-hosted gets everything managed does.

Agent isolation

Each agent runs in its own OCI container. Resource limits enforced per invocation. No shared process state.

Audit trail

Every invocation is a durable workflow — the exact inputs, outputs, tool calls, LLM calls, reviewer actions are preserved and replayable.

Ready to see it running?

Book a 45-minute walkthrough — the portal, the human-in-the-loop inbox, a working customer-service agent on real retail data.

Book a walkthroughReference agents