The best way to control Claude Code, Codex and more from any computer or phone
| 1 | --- |
| 2 | title: Vertex AI |
| 3 | description: Call Gemini on Vertex AI from Cloud Run with the GenerateContent binding — no API key, just the service's own identity. |
| 4 | --- |
| 5 | |
| 6 | `GCP.AIPlatform.GenerateContent` binds a Gemini model to your service. |
| 7 | The service calls Vertex AI as itself, so there's no API key to store |
| 8 | or rotate. |
| 9 | |
| 10 | ## Bind a model |
| 11 | |
| 12 | ```typescript |
| 13 | // src/Chat.ts |
| 14 | import * as GCP from "@alchemy.run/gcp"; |
| 15 | import * as Effect from "effect/Effect"; |
| 16 | import { HttpServerRequest } from "effect/http/HttpServerRequest"; |
| 17 | import * as HttpServerResponse from "effect/http/HttpServerResponse"; |
| 18 | |
| 19 | export default class Chat extends GCP.Function<Chat>()( |
| 20 | "Chat", |
| 21 | { main: import.meta.url }, |
| 22 | Effect.gen(function* () { |
| 23 | const gemini = yield* GCP.AIPlatform.GenerateContent("gemini-2.5-flash"); |
| 24 | |
| 25 | return { |
| 26 | fetch: Effect.gen(function* () { |
| 27 | const request = yield* HttpServerRequest; |
| 28 | const { prompt } = (yield* request.json) as { prompt: string }; |
| 29 | |
| 30 | const response = yield* gemini.generate({ |
| 31 | contents: [{ role: "user", parts: [{ text: prompt }] }], |
| 32 | }); |
| 33 | |
| 34 | const text = (response.candidates?.[0]?.content?.parts ?? []) |
| 35 | .filter((part) => !part.thought) |
| 36 | .map((part) => part.text ?? "") |
| 37 | .join(""); |
| 38 | return yield* HttpServerResponse.json({ text }); |
| 39 | }).pipe(Effect.orDie), |
| 40 | }; |
| 41 | }).pipe(Effect.provide(GCP.AIPlatform.GenerateContentHttp)), |
| 42 | ) {} |
| 43 | ``` |
| 44 | |
| 45 | The binding grants the service `roles/aiplatform.user`. Gemini 2.5 can |
| 46 | return "thought" parts, so filter them out of the answer. |
| 47 | |
| 48 | ## Pick a region |
| 49 | |
| 50 | The model runs on the `global` endpoint by default, which sends each |
| 51 | request wherever there's capacity. Pin a region instead: |
| 52 | |
| 53 | ```typescript |
| 54 | const gemini = yield* GCP.AIPlatform.GenerateContent({ |
| 55 | model: "gemini-2.5-flash", |
| 56 | location: "us-central1", |
| 57 | }); |
| 58 | ``` |
| 59 | |
| 60 | ## Tune a request |
| 61 | |
| 62 | System instructions and generation settings go on each call: |
| 63 | |
| 64 | ```typescript |
| 65 | const response = yield* gemini.generate({ |
| 66 | contents: [{ role: "user", parts: [{ text: prompt }] }], |
| 67 | systemInstruction: { parts: [{ text: "Answer in one sentence." }] }, |
| 68 | generationConfig: { temperature: 0.2, maxOutputTokens: 256 }, |
| 69 | }); |
| 70 | ``` |
| 71 | |
| 72 | ## Handle rejected prompts |
| 73 | |
| 74 | A blocked prompt or a bad setting fails with `BadRequest`. Return it to |
| 75 | the caller instead of a bare 500: |
| 76 | |
| 77 | ```typescript |
| 78 | .pipe( |
| 79 | Effect.catchTag("BadRequest", (error) => |
| 80 | HttpServerResponse.json({ error: error.message }, { status: 400 }), |
| 81 | ), |
| 82 | Effect.orDie, |
| 83 | ) |
| 84 | ``` |
| 85 | |
| 86 | ## Keep it private |
| 87 | |
| 88 | A public chat endpoint spends money on anyone's prompts. Leave |
| 89 | `invokerIamDisabled` off and call it from your own services with |
| 90 | [`InvokeService`](/gcp/networking/private-services). |
| 91 | |
| 92 | ## Example |
| 93 | |
| 94 | [`examples/gcp-vertex-ai`](https://github.com/alchemy-run/alchemy/tree/main/examples/gcp-vertex-ai) |
| 95 | is a chat endpoint with a live test. |
| 96 | |
| 97 | ## Where next |
| 98 | |
| 99 | - [Cloud Run](/gcp/compute/cloud-run) — the service the model is bound |
| 100 | to. |
| 101 | - [AIPlatform reference](/providers/gcp/aiplatform#generatecontent) — |
| 102 | every Vertex AI resource and binding. |