# pm-aigateway

Self-hosted AI gateway powered by [agentgateway](https://agentgateway.dev) — a unified, OpenAI-compatible proxy for LLM providers, MCP servers, and A2A agents.

## Project goal

Provide a single API endpoint that routes LLM requests to multiple providers (OpenAI, Anthropic, etc.) with centralized auth and observability. Stateless to start; extend to MCP and A2A as needed.

## Deployment targets

| Target | Status |
|--------|--------|
| Docker Compose | **current** — single stateless container |
| Kubernetes | planned |

## Stack

- **agentgateway** — the sole service; stateless, single binary in a container

No database or cache needed for the current scope.

## Repository layout

```
.
├── CLAUDE.md
├── compose/
│   ├── docker-compose.yml
│   └── .env.example
└── config/
    └── config.yaml          # model list and routing rules
```

## Key configuration

All provider API keys are supplied via environment variables (never committed).  
Copy `compose/.env.example` → `compose/.env` and fill in values.

Required env vars:
- `OPENAI_API_KEY`
- `ANTHROPIC_API_KEY`

agentgateway reads `config/config.yaml` on startup. Restart the container to pick up changes.

## Common commands

```bash
# Start
docker compose -f compose/docker-compose.yml --env-file compose/.env up -d

# Tail logs
docker compose -f compose/docker-compose.yml logs -f

# Restart after config change
docker compose -f compose/docker-compose.yml restart

# Stop
docker compose -f compose/docker-compose.yml down
```

## Ports

| Endpoint | Port |
|----------|------|
| OpenAI-compatible API | 4000 |
| agentgateway UI | 15000 |

## Test

```bash
curl -s http://localhost:4000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-4o-mini",
    "messages": [{"role": "user", "content": "Hello"}]
  }' | jq .
```

## Coding conventions

- Secrets in `compose/.env` only — gitignored. `compose/.env.example` is committed.
- Pin the agentgateway image tag; avoid `latest` in production.
- Keep `config/config.yaml` as the single source of truth for model/routing config.

## Future extensions

- **MCP gateway**: add `mcp.servers` block to `config.yaml` to expose tools
- **A2A gateway**: add `a2a` block for agent-to-agent routing
- **Persistence**: add Postgres if spend tracking or virtual keys are needed
- **Kubernetes**: translate compose to k8s manifests under `k8s/`