Ploos AI Gateway

A lightweight, self-hostable, provider-neutral AI API gateway built in Go.

Status: Active development. Authenticated chat routing and initial image generation are implemented. Not yet production-ready.

Principles

• Open source: MIT-licensed, self-hostable, no mandatory Ploos services.
• Provider-neutral: adapters for direct APIs and model aggregators.
• Privacy-conscious: no prompt or response persistence by default.
• Small footprint: single Go binary, standard library first.
• Safe defaults: listen on loopback by default; model route requires an API key.
• Portable: separate core logic from infrastructure and managed hosting.

Current capabilities (through PR #10)

• Authenticated OpenAI-compatible POST /v1/chat/completions with SSE streaming.
• Explicit JSON model aliases for OpenRouter, Ollama and private-IP self-hosted chat servers.
• Safe upstream error handling, timeouts, request limits and integration tests.
• Multimodal operation descriptors for chat, image, audio and video (metadata, not live endpoints).
• Authenticated POST /v1/images/generations using a configured OpenAI image model.
• Optional image_providers and image_models in the shared JSON registry.
• No implicit local-to-remote fallback.
• ComfyUI is not yet operational: its native workflow adapter remains unimplemented; configurations fail closed.

Roadmap

Phase	Status	Scope
PRs #1–#3	Implemented	Go service, CI, authentication, chat API and initial providers
PRs #4–#7	Implemented	Validation, private chat routing and integration tests
PR #8	Implemented	Multimodal capability contracts
PR #9	Implemented	Initial OpenAI text-to-image API
PR #10	Implemented	Shared image registry and fail-closed routing
Native ComfyUI	Next	Workflow templates, job polling and safe image retrieval
Image hardening	Planned	More providers, quotas and observability
Audio	Planned	Speech synthesis and transcription
Video	Planned	Async jobs; optional independent Ploos AI Video Connector integration
Operations	Planned	API key lifecycle, usage, budgets, rate limiting
Additional LLMs	Planned	Anthropic/Gemini adapters and explicit routing policies

Local-only model requests must never silently fall back to remote providers.

See architecture, image generation, image routing, and multimodal contracts.

Quick start

Requires Go 1.23 or newer.

export PLOOS_GATEWAY_API_KEY="replace-with-a-long-random-secret"
go run ./cmd/gateway
curl --fail http://127.0.0.1:8080/healthz

To enable inference, configure PLOOS_GATEWAY_REGISTRY and appropriate upstream credentials. See model routing and providers. Without a configured provider or registry, inference is unavailable.

Override the address with PLOOS_GATEWAY_LISTEN (for example, 127.0.0.1:9090).

Docker Compose

docker compose up --build -d
curl --fail http://127.0.0.1:8080/healthz
docker compose down

The development container is published only on localhost. Place a properly secured reverse proxy in front of a later production release; remember that Docker-published ports can bypass host firewall expectations.

Development

go fmt ./...
go vet ./...
go test ./...
go build ./cmd/gateway

See CONTRIBUTING.md and SECURITY.md.

Copyright (c) 2026 Ploos AS. Licensed under the MIT License.