01.1 Solo
SQLite, local files, embedded work.
The default profile serves a single-node installation with SQLite persistence, local blob storage, and an embedded durable worker—the direct path for personal and small deployments.
LIBRE / PRODUCT Deployment and operations CURRENT CAPABILITIES
Platform
Start on one machine. Grow deliberately.
Libre keeps the default local-first path small, then introduces shared infrastructure as an explicit team profile. Costs, quality, logs, telemetry, backups, and failover evidence become first-class parts of operating the system.
- Explicit runtime profiles
- Solo + team
- OpenTelemetry export
- Opt-in
- Release failover drill
- 3 replicas
01 Profiles
Use the infrastructure the deployment actually needs.
Solo and team are coherent operating profiles, not a hidden continuum. Startup validates the selected shape so half-configured shared state does not quietly split across backends.
01.2 Team
Shared state across replicas and workers.
The team profile uses PostgreSQL, Redis coordination, S3-compatible encrypted blob storage, vector search, and external durable workers, with Docker Compose and Helm deployment paths.
01.3 Migration
Move from solo through a packaged workflow.
A migration CLI inventories and copies an existing SQLite installation into the team profile with checkpoints and resumable application steps.
02 Durability + availability
Treat work as recoverable state, not a lucky connection.
Long-running chat and media operations use durable jobs and replayable events. The team profile extends that model across application replicas and external workers.
02.1 Durable work
Resume streams and recover accepted jobs.
Responses are recorded while they stream, cancellation reaches the upstream request, and accepted work can recover after a process failure instead of vanishing with one connection.
02.2 Release-tested HA
Exercise the real three-replica shape.
The release gate runs a three-replica drill covering stream resumption, worker failure, Redis outage fallback, revocation, rate limits, object deletion, and coordinated scheduling.
02.3 Helm controls
Reference secrets and narrow network paths.
The chart can reference an existing Secret and optionally install NetworkPolicies for the application and workers. Policy enforcement still depends on the cluster CNI.
03 Cost + quality
Measure what models cost and how they perform.
Operations are not only uptime. Libre gives administrators tools to understand provider-reported usage, control spend, and compare output quality with repeatable evidence.
03.1 Tariffs
Version pricing instead of rewriting history.
Per-model input and output rates take effect at a chosen time. Historical usage keeps the tariff revision that priced it, and unreported provider usage is not invented.
Tariffs 03.2 Budgets
Warn on the way and stop at the limit.
Budgets can observe, alert, or block. Alert-and-block mode rejects new interactive generations when the briefly cached spend reaches the limit; scheduled automations remain metered but are not blocked.
Budgets 03.3 Evaluations
Combine feedback, blind battles, and saved runs.
Message feedback captures ratings and topics, the arena computes an Elo leaderboard from blind votes, and saved sets make evaluation runs reproducible.
Evaluations 04 Operate + recover
Observe without silently exporting.
Logs, telemetry, security records, and recovery tools are distinct surfaces with explicit data boundaries.
04.1 Logs + telemetry
Structured locally; exported only when configured.
JSON logs include correlation identifiers and secret redaction. Setting an OTLP/HTTP endpoint opts into OpenTelemetry spans for requests and durable jobs; leaving it unset exports nothing.
Observability settings 04.2 Security
Sessions, scopes, SSO, groups, and an audit trail.
Revocable sessions, hashed scoped tokens, generic OIDC, live group membership, centralized authorization, and a redacted append-only security audit log form the account boundary.
Authentication and security 04.3 Recovery
Inventory, back up, verify, and restore.
Packaged maintenance commands create signed, verifiable backups for solo and team profiles. A recovery inventory reports the state present and blockers before a snapshot starts.
Recovery readiness