Gemini Web2API - Vercel Deployment Documentation

中文文档 | Cloudflare Docs | Netlify Docs | Deno Docs Run Gemini Web2API on Vercel Edge Functions for SSE streaming — zero config, no server to maintain.

The endpoint lives at api/gemini.js, a thin Vercel adapter over the shared core in src/ (the same core used by the Cloudflare worker and the Netlify functions). vercel.json routes every path (e.g. /v1/chat/completions, /health) to this single function via rewrites, and the export const config = { runtime: 'edge' } at the bottom of the file runs it on the Edge runtime. Why vercel.json + public/? This repo is API-only (no static frontend). Vercel’s “Other” framework preset expects an output directory named public/ — without it the build fails with No Output Directory named "public" found. The empty public/.gitkeep satisfies that, and the rewrite rules make the function reachable at the same paths as the Netlify/Cloudflare deployments.

⚡ Quick Deploy

Option 1: 1-Click Deploy to Vercel

Click the Deploy button above.
Connect your GitHub account and choose a repository name.
(Optional) In Environment Variables, configure API_KEY or COOKIE_STRING.
Click Deploy. Your API is live at https://your-app-name.vercel.app.

Option 2: Connect via Vercel Dashboard (Git)

Fork or push this repository to your GitHub account.
Open the Vercel Dashboard.
Click Add New… → Project and import your repository.
In Configure Project:
- Framework Preset: Other
- Build Command: (leave empty)
- Output Directory: (leave empty)
- Install Command: (leave empty)
Click Deploy. The repo’s vercel.json sets the output directory (public/) and rewrites all paths to api/gemini — no build settings needed.

Option 3: Vercel CLI

npm install -g vercel
vercel          # preview deployment
vercel --prod   # production deployment

⚙️ Environment Variables (Optional)

Configure these in Vercel: Project Settings → Environment Variables (or vercel env add).

Variable	Type	Description	Default
`API_KEY` / `API_KEYS`	String	Authorized client API keys (comma or `\|` separated).	`sk-gemini`
`COOKIE_STRING`	String	Google account cookies (`__Secure-1PSID`, etc.) for authenticated access.	`null`
`SAPISID`	String	SAPISID value (auto-extracted from `COOKIE_STRING` if omitted).	`null`
`DEFAULT_MODEL`	String	Default model if client does not specify one.	`gemini-3.6-flash`
`GEMINI_BL`	String	Gemini web build label (e.g. `boq_assistant-bard-web-server_...`).	built-in latest
`RATE_LIMIT_MAX`	Number	Max requests per IP in the window.	`3000`
`RATE_LIMIT_WINDOW`	Number	Rate limit window in seconds.	`60`
`REQUEST_TIMEOUT_SEC`	Number	Per-attempt upstream timeout in seconds. Keep small (e.g. `10`) to stay inside the 25s header deadline.	`28`
`ENABLE_PROXY`	String	Enables the rotating proxy pool on platforms with raw TCP sockets (e.g. Cloudflare Workers). Not supported on Vercel Edge (no TCP socket API) — setting it has no effect; a `WARN` is logged at request time.	`false`
`HTTPS_PROXY`	String	Static outbound proxy URL. Informational only on Vercel: unlike Netlify Edge (Deno), Vercel’s `fetch()` does not automatically tunnel through `HTTPS_PROXY`, so this value is reported by `/health` but not applied.	none (direct)

Proxy support on Vercel: Vercel Edge Functions do not provide raw TCP sockets (cloudflare:sockets), so the rotating proxy pool (ENABLE_PROXY/PROXY_ENABLED) is not supported on Vercel and is ignored with a WARN in the logs. Unlike Netlify Edge (Deno-based, native HTTPS_PROXY tunneling), Vercel’s Edge runtime does not auto-tunnel fetch() through HTTPS_PROXY either — direct connection is the only outbound mode on Vercel. If you need a proxy, use the Cloudflare Workers deployment. The /health endpoint reports the effective mode under proxy.mode (direct on Vercel).

Tip for Multiple Cookies: You can rotate between multiple Google accounts by separating cookies with a pipe character (|), e.g. cookie_account_1| cookie_account_2.

Platform Limits (Edge Runtime)

Limit	Value	How this deployment handles it
First response byte	Must start within 25s	Non-streaming requests cap all retries at a 22s pre-response deadline (3s safety margin). Streaming responses send headers immediately, so they are unaffected.
Total streaming duration	Up to 300s	Plenty for SSE token streaming.
Request/response body	4.5 MB	Long conversations may hit this; keep payloads reasonable.
Bundle size	250 MB uncompressed	The handler is a single ~155 KB file — no concern.

Verification

Once deployed, check your health endpoint in your browser or with curl:

curl https://your-app-name.vercel.app/health

Expected response:

{
  "status": "ok",
  "version": "1.7.3-multi-platform",
  "platform": "Vercel Edge Functions",
  "models": [
    "gemini-3.8-flash",
    "gemini-3.7-flash",
    "gemini-3.6-flash",
    "gemini-3.5-flash-thinking",
    "gemini-3.1-pro",
    "gemini-auto",
    ...
  ],
  "defaultModel": "gemini-3.6-flash",
  "hasCookie": false,
  "hasSapisid": false,
  "proxy": {
    "mode": "direct",
    "enabled": false,
    "poolSupported": false,
    "outboundProxy": null
  }
}

On Vercel, poolSupported is always false and proxy.mode is always direct — the platform offers no TCP sockets and no automatic proxy tunneling.

Client Configuration

NextChat / ChatGPT-Next-Web

Field	Value
Interface Type	OpenAI
Endpoint / Base URL	`https://your-app-name.vercel.app/v1`
API Key	`sk-gemini` (or your configured `API_KEY`)
Model	`gemini-3.6-flash` or `gemini-3.5-flash-thinking`

Cherry Studio / ChatBox

Field	Value
Provider	OpenAI
API Base URL	`https://your-app-name.vercel.app/v1`
API Key	`sk-gemini`
Model	`gemini-3.6-flash`

curl Test

curl https://your-app-name.vercel.app/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer sk-gemini" \
  -d '{
    "model": "gemini-3.6-flash",
    "stream": true,
    "messages": [{"role": "user", "content": "Hello Gemini!"}]
  }'

🛠️ Local Development

You can run and test the Vercel Edge Function locally using the Vercel CLI:

npm install -g vercel
vercel login   # first time only
vercel dev

The local server will start at http://localhost:3000, with the same rewrite routing as production. (The CLI requires a free Vercel account even for local dev.)

Troubleshooting

`No Output Directory named "public" found after the Build completed`

This repo ships a vercel.json with "outputDirectory": "public" and a public/.gitkeep placeholder, so this should not happen on a fresh deploy. If you hit it anyway:

Make sure vercel.json and public/.gitkeep are committed and pushed.
Or set Output Directory to public under Project Settings → Build & Output Settings.

`FUNCTION_INVOCATION_TIMEOUT` (504)

Vercel Edge Functions must begin sending a response within 25 seconds, otherwise the invocation is terminated with a 504. Non-streaming requests now cap all retries at a 22s pre-response deadline. If you still see this:

Cause	Fix
Gemini upstream unreachable or slow	Lower `REQUEST_TIMEOUT_SEC` (e.g. `10`) so a single attempt cannot consume the deadline.
`GEMINI_BL` build label expired (upstream returns 405)	Open `https://gemini.google.com/app`, press F12 → Network tab, search any request URL for `boq_assistant`, copy the newest label into the `GEMINI_BL` environment variable.
Rate limited by Google (upstream returns 429)	Add a valid `COOKIE_STRING` (+ `SAPISID`) environment variable, or lower request frequency.

`FUNCTION_INVOCATION_FAILED` (500)

The function crashed with an unhandled exception. The handler has a top-level try/catch that returns a structured JSON 500, so this usually means the error happened before the handler was invoked. Check Deployments → Runtime Logs in the Vercel dashboard.

Malformed request (non-string `model`, non-object JSON body)

Fixed: invalid input returns a structured 400 instead of crashing.

Logs

Runtime logs are available in the Vercel dashboard under Deployments → Runtime Logs (or vercel logs <deployment-url>). The handler logs requests as [HH:MM:SS] [LEVEL] message when logRequests is enabled (default).

Getting Started

Deployment

Additional Info

Free gemini vercel

Gemini Web2API - Vercel Deployment Documentation

⚡ Quick Deploy

Option 1: 1-Click Deploy to Vercel

Option 2: Connect via Vercel Dashboard (Git)

Option 3: Vercel CLI

⚙️ Environment Variables (Optional)

Platform Limits (Edge Runtime)

Verification

Client Configuration

NextChat / ChatGPT-Next-Web

Cherry Studio / ChatBox

curl Test

🛠️ Local Development

Troubleshooting

`No Output Directory named "public" found after the Build completed`

`FUNCTION_INVOCATION_TIMEOUT` (504)

`FUNCTION_INVOCATION_FAILED` (500)

Malformed request (non-string `model`, non-object JSON body)

Logs

​Gemini Web2API - Vercel Deployment Documentation

​⚡ Quick Deploy

​Option 1: 1-Click Deploy to Vercel

​Option 2: Connect via Vercel Dashboard (Git)

​Option 3: Vercel CLI

​⚙️ Environment Variables (Optional)

​Platform Limits (Edge Runtime)

​Verification

​Client Configuration

​NextChat / ChatGPT-Next-Web

​Cherry Studio / ChatBox

​curl Test

​🛠️ Local Development

​Troubleshooting

​No Output Directory named "public" found after the Build completed

​FUNCTION_INVOCATION_TIMEOUT (504)

​FUNCTION_INVOCATION_FAILED (500)

​Malformed request (non-string model, non-object JSON body)

​Logs

Gemini Web2API - Vercel Deployment Documentation

⚡ Quick Deploy

Option 1: 1-Click Deploy to Vercel

Option 2: Connect via Vercel Dashboard (Git)

Option 3: Vercel CLI

⚙️ Environment Variables (Optional)

Platform Limits (Edge Runtime)

Verification

Client Configuration

NextChat / ChatGPT-Next-Web

Cherry Studio / ChatBox

curl Test

🛠️ Local Development

Troubleshooting

`No Output Directory named "public" found after the Build completed`

`FUNCTION_INVOCATION_TIMEOUT` (504)

`FUNCTION_INVOCATION_FAILED` (500)

Malformed request (non-string `model`, non-object JSON body)

Logs