on-premise deployment

the whole platform.inside your perimeter.

Not an API you call — the entire Donkit platform, deployed in your private cloud or on your own servers. Your agents, your data, your models. Nothing leaves your network.

01 — why on-premise

Most "enterprise AI" means sending your data to someone else's cloud and trusting the paperwork. For some teams that's fine. For legal, finance, healthcare, defense, and anyone with strict data-residency rules, it's a non-starter.

Donkit takes the opposite path: the platform comes to your data. Agents, orchestration, evaluation, vector search, storage — everything runs where you decide.

02 — what ships

the full stack.
not a lite edition.

One deployment brings the complete product — the same one that runs Donkit's cloud, not a stripped-down "enterprise edition".

component · 01
01

control plane

Agent management, the builder, and evaluation orchestration — the brain of the platform.

component · 02
02

agent runtime

Where your agents execute, with tool and MCP access to your internal systems.

component · 03
03

background workers

Document ingestion, evaluation runs, and the self-optimization loop — always on.

component · 04
04

data layer

PostgreSQL, Qdrant vector store, message queue, and S3-compatible object storage.

Everything you see in the SaaS — the meta-agent builder, interventional evals, generated UIs, per-agent databases, token-level tracing — works identically on-premise.

03 — kubernetes-native

one helm chart.
your cluster.

Deployed with a single Helm chart into your Kubernetes cluster — AKS, EKS, GKE, or bare-metal. GitOps-friendly: manage it with ArgoCD like the rest of your infrastructure.

For select customers, we offer full-scale rollout pilots with special conditions — reach out to see if you qualify.

$ helm repo add donkit https://charts.donkit.ai
$ helm install donkit donkit/platform -f values.yaml
# control-plane ✓ runtime ✓ workers ✓ data ✓
STATUS: deployed — 0 egress connections
AKSEKSGKEbare-metalArgoCD-ready
# llm-endpoints.yaml
providers:
  - azure_openai: https://your-tenant.openai.azure.com
  - vertex_ai:    https://your-region-aiplatform.googleapis.com
routing: auto — per-request complexity · domain · cost
Azure OpenAIGoogle Vertex AIOpenRouterFireworksyour local LLM
04 — your models, your rules

bring the endpoints
you already trust.

Donkit doesn't lock you into a model vendor. Point the platform at the LLM endpoints your organization has already approved, including local LLM deployments, and Donkit's auto-routing picks the right one per request.

Prompts, documents, and agent data never cross your network boundary.

05 — security

security by
architecture.

Not security by policy document — the isolation is structural, and it inherits the controls you already run.

secrets

Encrypted at rest

Integration credentials and connection secrets are encrypted, decrypted only inside the runtime that uses them.

isolation

Per-agent data isolation

Each agent's database is separate, accessed under a restricted role that physically cannot drop or destroy data.

audit

Full audit trail

Every SQL query an agent runs, every schema change, is logged with role, duration, and result.

controls

Your IAM, your policies

The platform runs inside your cluster and inherits your network policies and access controls.

06 — rollout

how teams
roll it out.

From first install to production in three phases — no big-bang migration.

phase · 01
01

pilot.

Stand up the stack, connect a test model endpoint, build the first agent the same day.

phase · 02
02

integrate.

Wire your internal systems via MCP, load knowledge bases, set up SSO and network policy.

phase · 03
03

production.

GitOps-managed upgrades; new platform versions arrive as chart updates you apply on your schedule.

07 — faq

01Is it feature-equivalent to the cloud?

Yes — same codebase, same capabilities.

02What do we need to run it?

A Kubernetes cluster and access to an LLM endpoint you trust. Sizing depends on workload, but enterprise customers comfortably run the platform on 6-core systems with 20 GB RAM and 50 GB SSD.

03How do updates work?

Versioned Helm chart releases; you control when to apply.

04Can it run fully air-gapped?

Yes. Point the platform at an LLM endpoint inside the perimeter — including a local model — and nothing crosses the boundary.

05Pricing?

Enterprise pricing — talk to us.

ready · 08

bring the agents
to your data.

book a demo