Skip to content

Releases: IamNotShady/LLMIngress

LLMIngress v0.1.0-beta.2

Pre-release

Choose a tag to compare

@IamNotShady IamNotShady released this 30 Jul 03:04
54077a2

Second beta of LLMIngress — a self-hosted LLM gateway and console.

Highlights

  • 22 new provider templates (13 → 35 built-in)
    • Subscription (OAuth): Grok and MiniMax Coding Plan join Claude Code and OpenAI Codex. MiniMax Coding Plan introduces the device-code flow: show a user code, open the verification page, poll to completion — no redirect URI needed
    • API key: Anthropic, AWS Bedrock, BytePlus ModelArk, Cerebras, ClinePass, Command Code, Fireworks AI, GLM Coding Plan, Groq, Kimi Coding Plan, Mistral, Mistral Vibe, NousResearch, NVIDIA NIM, Ollama Cloud, OpenAI, OpenCode Go, Qwen Token Plan, Xiaomi MiMo, Xiaomi MiMo Token Plan
  • Upstream quota tracking: providers that report subscription windows, monthly budgets, or token plans get a periodic Worker quota probe; the Console shows remaining quota per connection
  • Gateway circuit breaker: per-connection retries with exponential backoff, plus an error-rate breaker (sliding window, half-open recovery) that short-circuits failing connections before they burn requests
  • Cross-catalog model metadata fallback: models missing capability or price data in one catalog source now fall back to the others (models.dev, OpenRouter, LiteLLM, Vercel), with expanded price-sync coverage and prefix-vendor resolution
  • Console UI redesign: refreshed layout and theming across all pages, dialogs, and drawers
  • Deploy improvements: branch-scoped Compose projects (isolated containers, network, and PostgreSQL volume per branch), automatic repair of stale container-network state, and one unified env-file precedence (shell → .env.local.env → defaults) across every startup path

Deployment

Two-container Docker Compose delivery (one app container + PostgreSQL 18.4):

git clone https://github.com/IamNotShady/LLMIngress.git
cd LLMIngress
git checkout v0.1.0-beta.2
./scripts/deploy.sh

deploy.sh generates a random ENCRYPTION_KEY into .env when missing, then starts Compose. The Console listens on http://127.0.0.1:3000 and the Gateway on http://127.0.0.1:4000; all ports bind to loopback only by default.

⚠️ Back up the ENCRYPTION_KEY in .env — stored provider credentials can only be decrypted with the same key.

Upgrading from v0.1.0-beta.1

Check out the new tag and re-run ./scripts/deploy.sh with the same .env. The app container applies the two new migrations (0002_provider_quota, 0003_provider_oauth_device) automatically on start. Keep the same ENCRYPTION_KEY.

Note: on the main branch the Compose project name is unchanged (llmingress), so existing containers and the PostgreSQL volume carry over.

Known limitations (beta)

  • PostgreSQL uses the Compose-bundled password and its port is published to 127.0.0.1 only; adjust the Compose file if you need external access
  • /v1/embeddings is not supported in this release

LLMIngress v0.1.0-beta.1

Pre-release

Choose a tag to compare

@IamNotShady IamNotShady released this 18 Jul 15:13
2e6e294

First public beta of LLMIngress — a self-hosted LLM gateway and console.

Highlights

  • Gateway: OpenAI Chat Completions / Responses and Anthropic Messages protocols with streaming, safe fallback chains, and request logging
  • Virtual Model routing: atomic Virtual Models with fixed, cost-first, and load-balance ordering plus fallback chains
  • Provider management: API key and OAuth credentials (encrypted at rest with ENCRYPTION_KEY), model catalog sync, exact connection health probes; model lists refresh automatically once credentials are ready
  • API Key access control: API keys, Virtual Model grants, and budget / rate / token / concurrency limits
  • Usage & activity: tokens, latency, failures, fallbacks, and actual or estimated cost
  • Console: eight pages — Overview, API Keys, Virtual Models, Providers, Activity, Usage, Limits, Playground
  • Worker: durable model_refresh, provider_connection_probe, and price_sync jobs with a self-healing poll loop and timeout-bounded network calls

Deployment

Two-container Docker Compose delivery (one app container + PostgreSQL 18.4):

git clone https://github.com/IamNotShady/LLMIngress.git
cd LLMIngress
git checkout v0.1.0-beta.1
./scripts/deploy.sh

deploy.sh generates a random ENCRYPTION_KEY into .env when missing, then starts Compose. The Console listens on http://127.0.0.1:3000 and the Gateway on http://127.0.0.1:4000; all ports bind to loopback only by default.

⚠️ Back up the ENCRYPTION_KEY in .env — stored provider credentials can only be decrypted with the same key.

Known limitations (beta)

  • PostgreSQL uses the Compose-bundled password and its port is published to 127.0.0.1 only; adjust the Compose file if you need external access
  • /v1/embeddings is not supported in this release

Note: this tag was re-cut on 2026-07-18 and now points at a newer commit than it did on 2026-07-16. If you cloned or fetched v0.1.0-beta.1 before that date, git will not update the tag on a plain git fetch — run git fetch --tags --force to pick up the current code. The rebuilt beta renames the Agent entity to API Key, renames MASTER_KEY to ENCRYPTION_KEY, and renames the random route policy to load_balance.