DocsSelf-hosted (Enterprise)

Self-hosted EvaliQA

Run EvaliQA inside your own infrastructure on the Custom plan: private-network targets, local models, several workspaces and your own platform model.

Last updated 2026-10-14

Self-hosted EvaliQA runs inside your own infrastructure instead of on our hosted service. Your projects, datasets, runs and traces stay in your network, on your own databases and object storage. The EvaliQA team delivers the installation and its updates.

Self-hosting is part of the Custom plan, licensed per environment. See Pricing and talk to us there.

What stays the same

The product is the same: projects, connectors, test plans, datasets, eval runs, reports, tracing, online evaluation and alerts work as described in the rest of these docs. Data retention is set for your installation rather than by plan.

What is different from the hosted service

Connectors and alerts can reach your private network

On the hosted service, connectors and alert channels only call public internet addresses. localhost, internal host names and private ranges such as 10.x, 172.16.x to 172.31.x and 192.168.x are refused, so a system on your LAN cannot be tested and a webhook on your LAN cannot be notified.

On a self-hosted installation your operators can allow private addresses, for all targets or only for the hosts you list. Then a connector can call an AI system that never leaves your network, and Slack-style webhooks and generic webhooks can post to internal receivers.

Local and self-hosted models

Credentials can point at models running in your network: Ollama, vLLM or any OpenAI-compatible endpoint (use the Custom provider for those). EvaliQA calls the model from its own servers, so the address must be reachable from where EvaliQA runs, not only from your laptop. See For your operators.

Several workspaces per person

On the hosted service a person belongs to one workspace. On a self-hosted installation one person can belong to several, for example one per team or per client, and switch between them.

Your own platform AI model

The platform AI model (AI hints, and, when the workspace uses platform credits, the judge, dataset generation and the simulated user) can be a model you choose, including one served inside your network. Workspaces can still set the Platform AI agent to their own credential in Settings → Platform AI agent.

For your operators

  • Connectors to private addresses. CONNECTOR_EGRESS_ALLOW_PRIVATE=true lets connectors call any address. To open only specific targets, leave it off and list them in CONNECTOR_EGRESS_ALLOWLIST, comma-separated: a host name (ai.corp.example), a wildcard (*.corp.example) or a URL prefix (http://10.0.4.12:8080/api/). Only http and https are ever allowed.
  • Alert channels to private addresses. ALERT_EGRESS_ALLOW_PRIVATE and ALERT_EGRESS_ALLOWLIST work the same way for Slack, Discord, Teams and webhook channels. Channels are checked when saved and again before every send.
  • Local models. Model credentials need no switch, only a reachable address. localhost means the EvaliQA container itself, not the machine it runs on. Use the model server's LAN address or DNS name (http://10.0.4.20:11434), or, when EvaliQA runs in Docker on the same machine as Ollama, http://host.docker.internal:11434 (available by default on Docker Desktop for Mac and Windows; on Linux map it to the host gateway).
  • Alerts. Channel secrets, deep links and email have their own settings, see Alerts, self-hosted notes.