@deslop/gateway 0.1.1250 → 0.1.1256

package/README.md

@@ -11,4 +11,16 @@ gateway service install --url https://gateway.mp281x.xyz --auth https://auth.mp2
 - **Admins** (`--admins`) add and connect servers; everyone signed in shares them.
 - **Own tool**: `list_mcp_connections` explains a missing or failing server.
 
+## AI proxy
+
+Each person manages their own Anthropic, OpenRouter, OpenCode Zen and OpenCode Go keys. Keys stay in Gateway's owner-only data files; only their last four characters and last upstream status leave the server.
+
+App servers start with `--gateway https://gateway.mp281x.xyz`. The runtime enrolls each authenticated person with a grant for that app's `--url`; enrollment forwards the session cookie or app-audience bearer to `POST /ai/grants`. A new grant replaces the previous grant for that person and app. Revoking it immediately stops new proxy calls. Without `--auth`, calls use the local developer and need no grant.
+
+Use `AiGateway` from `@deslop/runtime/ai`: `yield* gateway.apiUrl(provider)` and pass `gateway.authorize(userId)` as the provider client's **construction-time** `transformClient`. OpenRouter's scoped `withClientTransform` does not cover streaming chat completions. OpenRouter's base includes `/v1`; its Decisions client derives `/alpha/decisions` correctly. OpenCode bases expose protocol-family paths such as `/openai/v1/chat/completions` and `/anthropic/v1/messages` without translating them.
+
+`/ai/<provider>/<rest>` forwards method, query, body and upstream status. Fetch decodes compressed responses; stale encoding and length headers, hop-by-hop headers and headers nominated by `Connection` are removed from the streamed answer. Caller credentials and hop-by-hop headers are removed before inserting the person's provider key. Missing keys return HTTP 402 with instructions to add a key in Gateway. JSON and SSE usage is metered asynchronously; provider-reported cost wins, then daily-cached models.dev prices, otherwise cost remains unknown. Usage is stored per person and UTC month.
+
+Person-scoped RPCs: `ai.keys`, `ai.setKey`, `ai.removeKey`, `ai.grants`, `ai.revokeGrant`, `ai.usage`. Usage queries take epoch-millisecond `from` (inclusive), `to` (exclusive), and `bucket` (`hour` or `day`); `AiUsageReport` returns token/cost totals, UTC buckets split by app and model, and breakdowns by app, model and provider. `unknownCostCalls` distinguishes unknown cost from zero cost.
+
 Common flags and service commands: [Deslop app runtime](https://github.com/MP281X/deslop/blob/main/packages/runtime/README.md).