Models keep changing.
Capabilities, deployments and policies evolve. Applications need to know when their accepted expectations no longer hold.
Turn a deployed model endpoint into a governed AI service. Connect applications, shape access and limits, and make model changes explicit.
Self-hosted·Apache-2.0·One Docker image to start
chat · streaming · tools
fp_a91cTHE NEED
01 / CONTEXTA working endpoint is a starting point. Teams also need clear access rules, predictable limits and visibility into what happens when models or policies change.
Capabilities, deployments and policies evolve. Applications need to know when their accepted expectations no longer hold.
Teams, keys, quotas and guardrails need a shared place to configure, apply and inspect them.
A response should come with an explanation: selected model, attempts, checks, usage and estimated cost.
THE PLIQRA APPROACH
Applications can bind to a fingerprint of the service’s capabilities and gateway policy. A changed contract returns HTTP 409, so the application can deliberately rediscover and renew.
Protocol 0.1 is a draft. Binding is opt-in; standard supported inference APIs are also available.
01discoverchat, streaming, tools
02bindfingerprint: fp_a91c
03invoke409 ContractChanged
04renewExplicit application decision
TAKE THE CONTROLS
02 / INTERACTIVE DEMOEight short scenarios. Change a contract, exhaust a quota, inspect a guardrail, drain a site or review a fleet policy. No account or running model needed.
One workspace.
UI and API controls.
Discover a model, separate declared capabilities from verified behavior, and bind an application. Try changing the accepted contract.
Declared inventory and verified checks are tracked separately in Pliqra.
THE CAPABILITY MAP
03 / PRODUCTConfigure in the console. Automate through the API. Start with an endpoint and add the operational capabilities your workload needs.
Chat, embeddings, speech and transcription, Responses, Anthropic Messages and Ollama. Live native text/tool streams, within documented API subsets.
Read the guideService selection, project/team grants, request and token windows, budgets, concurrency, input/output bounds and scoped application keys.
Read the guideLocal PII, keyword and schema checks. NeMo check APIs and custom HTTP validators. Ordered input/output stages, enforce/observe modes and visible verdicts.
Read the guideOIDC, SAML, SCIM, workload JWT federation and project/team permissions. Environment, mounted-file and Vault secret references.
Read the guideCumulative administrator reporting, team and site filters, per-attempt routing, token accounting, unknown-usage handling and exports.
Read the guideImmutable prompt versions, durable batches, external judges, evaluations, quality gates and controlled canary traffic.
Read the guideOne writable PostgreSQL authority, strict shared admission, local-first placement, site drain, per-site health and safe fenced-process recovery.
Read the guideExplicit pairing, combined metadata visibility, resource mapping, second-administrator approval, rollout receipts and policy drift.
Read the guideDiscovery, capability/extension negotiation, fingerprinted bindings and deliberate renewal. Python, JavaScript/TypeScript and Go clients; draft protocol and executable core profile.
Read the guideEncrypted backups, recovery tools, monitoring, audit trails and retention controls. SQLite by default; PostgreSQL, organizations and analytics when needed.
Read the guideYOUR INFRASTRUCTURE
04 / TOPOLOGYPliqra runs alongside your already deployed models. A single installation starts with SQLite; shared deployments add PostgreSQL and your existing infrastructure.
Persist the data volume and encryption key. One SQLite process owns each volume. Configure HTTPS before exposing a workspace remotely.
Deployment detailsSHIPPED WITH EVIDENCE
Successful synthetic soak calls
Continuous synthetic acceptance
Observed synthetic p95 latency
2.1.0 acceptance · 7 November 2026. One host, 4 CPUs, PostgreSQL 18 and a deterministic 5 ms model fixture. Native AMD64/ARM64 smoke, model/database interruption recovery and RC4 SQLite upgrade/backup recovery also passed. Bounded regression evidence, not production sizing or an availability SLA.
Inspect the evidenceMAKE YOUR FIRST CONNECTION
Use the ready-made Linux AMD64/ARM64 image. Docker Desktop runs the matching Linux image on macOS or Windows. No source build or registry login required.
Docker with Compose is required. Run the commands, then open localhost:8088.
Create the administrator and add a reachable endpoint. Host-based models use host.docker.internal.
Configure the service and gateway, create a scoped key, then connect your application.
mkdir pliqra && cd pliqra
curl -fL -o compose.release.yaml https://github.com/eturkok/pliqra/releases/download/v2.1.0/compose.release.yaml
docker compose -f compose.release.yaml pull
docker compose -f compose.release.yaml up -dImage digests, checksums and upgrade guideA FEW CLEAR ANSWERS
MODELS WILL CHANGE. KEEP THE AGREEMENT VISIBLE.