AI Gateway API

Base URL: http://localhost:8088 (GATEWAY_HTTP_ADDR)

AI Gateway API

Source: apps/gateway/internal/server/server.go (route registration: Router())

Base URL: http://localhost:8088 (GATEWAY_HTTP_ADDR)

OpenAI-compatible, plus enterprise control-plane extensions.

Public routes

Method	Path	Auth	Description
GET	`/healthz`	None	Health check
POST	`/v1/chat/completions`	JWT Bearer	Chat (supports `stream: true` SSE)
POST	`/v1/embeddings`	JWT Bearer	Embeddings

Chat Completions

Headers

1Authorization: Bearer <portal_jwt>
2Content-Type: application/json

Optional routing headers (see routing.Decider):

Explicit provider header (config key provider_header)
local_route_header — force local / private-cloud / third-party

Body: standard OpenAI chat completions JSON.

Response

Non-streaming: OpenAI JSON + usage
Streaming: text/event-stream, SSE data lines
Policy block: business-error JSON (not upstream model refusal text), including the matched rule

Streaming hardening env vars

GATEWAY_STREAM_IDLE_TIMEOUT
GATEWAY_STREAM_SCANNER_MAX_BUFFER_MB

Internal routes

Method	Path	Auth	Description
GET	`/internal/channel-stats`	`GATEWAY_INTERNAL_TOKEN`	Channel runtime stats

Upstream key resolution

OpenAICompatibleProvider order:

<PROVIDER>_API_KEY (provider name uppercased, - → _)
LLM_API_KEY as a generic fallback
Unset → mock fallback (policy / audit / metering still run)

Keys configured in the Admin GUI are stored as a PG cipher. Gateway runtimeconfig polls about every ~5s; that source outranks env.

Config sources

Config	Env (remote)	Env (local)
Providers	`GATEWAY_REMOTE_PROVIDERS_URL`	`GATEWAY_ADMIN_PROVIDERS_FILE`
Quota	`GATEWAY_REMOTE_QUOTA_CONFIG_URL`	`GATEWAY_QUOTA_CONFIG_FILE`
Policy snapshot	`GATEWAY_REMOTE_POLICY_SNAPSHOT_URL`	`GATEWAY_POLICY_SNAPSHOT_FILE`
Channel	`GATEWAY_REMOTE_CHANNELS_URL`	PG `gateway_channels`
Policy override	—	`GATEWAY_POLICY_OVERRIDE_FILE`

Remote fetch requires GATEWAY_INTERNAL_TOKEN to match admin.

Audit and metering outputs

Output	Path / table
Audit JSONL	`apps/gateway/.runtime/audit/audit-*.jsonl`
Audit PG	`gateway_audit_events`
Pending backfill	`.runtime/audit/.pg-pending`
Metering PG	`usage_records`
Metering JSONL	`GATEWAY_USAGE_LOG` (when PG is absent)

Build and run

bash

1cd enterprise/apps/gateway
2go build -o bin/gateway ./cmd/gateway
3go run ./cmd/gateway

Docker:

bash

1cd enterprise
2docker build -f apps/gateway/Dockerfile -t agenticx-gateway:latest .

See ../gateway/overview.md and ../../apps/gateway/README.md.

Difference from Machi Desktop

Enterprise Gateway is a standalone Go process. Machi Desktop uses the embedded Python agx serve + LiteLLM. These are not the same implementation.

Made-with: Damon Li