Skip to main content
The gateway is a request handler. Run it as the self-hosted server, or compose it into your own service with a GatewayPlatform that supplies storage, accounting, and credentials. The reference platform is src/self_host/ in the gateway repository.

Serve The Handler

The root entry point uses only web-standard fetch, Request, and Response, so any runtime can serve it:
Set idleTimeout: 0 so long streams stay open while a model thinks. The Bun entry point (@mupt-ai/dari-gateway/bun) adds startServer for the same setup, and startSelfHostedGateway for the full self-hosted server with SQLite.

The Platform

A GatewayPlatform supplies:
  • Router resolution. Turns a router id and the caller’s gateway key into models, thinking levels, and settings.
  • Conversation state. Reads and writes the prefix state that keeps a conversation on one model and its provider caches warm.
  • Session pins. Reads and writes pins for routers with session_pinning: true, keyed by router and an opaque pin key.
  • Subscription holds. Usage-limit holds on connected plans, so a held plan is skipped until its reset.
  • Request records. Usage for each request: input, output, cache-read, and cache-write tokens, and total_cost_usd. Failed and aborted requests are recorded too.
  • Error reporting. Request and operational errors.

Credentials

Implement resolveProviderCredential to supply a per-request credential in place of the router’s own key. Return { apiKey, baseUrl?, headers?, subscription? }:
  • Pass a tenant’s own key for that request.
  • Set subscription to "openai_codex", "anthropic_claude_code", or "meta_muse" to serve from a plan. The apiKey is then that plan’s access token, and the platform keeps it fresh.
A Claude Code plan serves only Claude Code clients, and a Muse Code plan only Muse Code clients, recognized by their x-client-id.

Strict Routers

A router resolved with strict: true serves only the requested dari/model/<id>, at the requested thinking level. It uses no selector, lease, or personal subscription, and it refuses anything else instead of substituting a model. Set strict_providers: ["openai"] on the resolved router to allow any bundled-catalog model from that provider without listing each one. Use the full form, such as dari/model/openai/gpt-6.2-sol. Bare ids and models outside the catalog are still refused. The grant applies only to strict routers.

Session Pinning

With session_pinning: true, the runtime routes each session once and serves that model and level for the rest of the session. Each subagent is routed once on its own, identified by a Claude Code agent id, a Codex subagent thread, or an OpenCode child session. Background requests and dari/model/<id> requests never use pins.

Cost Attribution

A client can tag requests with X-Dari-Tags: run=r1,trial=3. The runtime parses the tags into the request record’s tags and into the recordUsage payload, so you can group spend by run or experiment.

Details

Each request record also carries the router’s non-secret settings, such as enabled models, thinking levels, and the strict, speculative-routing, and session-pinning flags.