v0.2.0

0.2.0 (2026-10-13)

⚠ BREAKING CHANGES

• remove edge nodes and the --mode flag

Features

• add a Rust data-plane gateway that joins gateway groups (17ab06d)
• add gateway groups and a sync protocol for data-plane gateways (3eee415)

Performance Improvements

• encode gateway usage batches off the request threads (a9f1472)

Code Refactoring

• remove edge nodes and the --mode flag (a902c36)

v0.1.2

0.1.2 (2026-10-07)

Features

• run workspaces from a containerised gateway (#8) (00b40e4)

v0.1.1

0.1.1 (2026-10-06)

Features

• optional Langfuse stack in the compose file (#3) (ec16b29)

Bug Fixes

• warn that pass-through models can't be used in the Playground (#2) (7626597)

v0.1.0

Initial release.

• OpenAI-compatible gateway (/v1/chat/completions, /v1/completions, /v1/embeddings,
  /v1/models) with streaming passthrough and exact token accounting
• Providers: Azure OpenAI / AI Foundry (Entra ID or api-key outbound auth, native price
  catalog), Google Vertex AI (Application Default Credentials or a service-account JSON key;
  Gemini and every Model Garden MaaS model through the
  OpenAI-compatible endpoint, Claude through rawPredict; publisher-catalogue discovery and
  gcp/ catalog prices), OpenRouter (discovery fills in its published prices from the
  catalog) and any OpenAI-compatible backend (vLLM, LM Studio, TGI, Ollama)
  with custom per-model pricing
• Pass-through auth mode (oauth_passthrough): forwards the caller's own upstream
  credential and anthropic-beta header instead of storing one, so Claude Code can run
  on a Claude subscription through the gateway (clients authenticate with X-Llmr-Key);
  that traffic is always recorded unpriced. LLMR_SEED_CLAUDE_SUBSCRIPTION=1 seeds it with
  one claude-* model, so new Claude models need no gateway update
• Wildcard model names: a model name ending in * serves every name with that prefix
  (exact names first, then the longest prefix), with the requested name sent upstream when
  the upstream name ends in * too
• Virtual API keys, users, teams; budgets (daily/monthly/total) and rpm/tpm rate limits
  per key/user/team; model policy on teams, users and keys (a person's default is the union of
  their teams' policies; keys only narrow), per-member budget shares within a team,
  member-minted team keys, and a Models page for every signed-in user
• Inbound Azure Entra ID / GCP identity-token authentication with claim mapping and
  optional auto-provisioning
• Web console: dashboards, key management, teams, providers/deployments, pricing,
  OIDC SSO, RBAC (admin / team admin / member)
• Hardening from the pre-release review: LLMR_ENCRYPTION_KEY is required and
  placeholder values are refused (it signs console sessions); the workspace IDE is served
  from its own origin (LLMR_IDE_LISTEN / LLMR_IDE_ORIGIN) and the proxy strips the
  caller's credentials before the container; console mutations are checked against the
  browser's Origin; keys of disabled users stop working; deleting a user removes their
  workspace containers and volumes; deleting a team asks before revoking its keys; the
  last enabled admin cannot be disabled or demoted and first-run setup only opens on an
  empty database; members cannot set budgets or rate limits on their own keys; database
  errors are no longer echoed to clients; streamed Claude usage counts cache writes
• Edge gateway mode: config long-poll sync, sealed disk cache for cold starts and
  control-plane outages, batched usage shipping
• Workspaces: per-user containers with VS Code (code-server), a terminal and a pi-powered
  chat; admin-managed templates (base, python, node, go, java, infra) with in-console
  Dockerfile editing and image builds; per-user files injected into every workspace;
  $HOME on the volume so toolchain caches persist; Docker/podman and Kubernetes runtimes
• Deploy: single static binary, Docker image, docker-compose, Helm chart (both modes)