Web2API

简体中文

Expose logged-in ChatGPT, Gemini, and Grok Chrome sessions through a local OpenAI-compatible Responses API.

Features

• Independent ChatGPT, Gemini, and Grok providers
• Text, streaming text, multi-turn conversations, and reasoning effort
• Image input from URLs or base64 data URLs
• Image generation and image editing
• Emulated function calling
• Local execution without storing prompts, responses, images, cookies, or provider credentials

Quick Start

Requirements: Google Chrome, Node.js 22+, and a logged-in ChatGPT, Gemini, or Grok account.

1. Install the ready-to-load extension. No repository clone or build is required:

npx -y glidea-web2api@latest prepare-extension

2. Open chrome://extensions, enable Developer mode, click Load unpacked, and select the Extension directory printed by the command.

3. Pin Web2API and open its popup. Run the one-time installation command shown there:

npx -y glidea-web2api@latest install --extension-id <extension-id>

4. Return to the popup and click Check again. Web2API starts the local companion and provider worker tabs. Sign in to the providers you need, then copy the Base URL and API key from the popup.

Verify the complete path:

export WEB2API_BASE_URL="http://127.0.0.1:3210/v1"
export WEB2API_KEY="wb2_replace_with_popup_value"

curl "$WEB2API_BASE_URL/responses" \
  -H "Authorization: Bearer $WEB2API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"chatgpt/default","input":"Reply with exactly WEB2API_OK"}'

Seeing WEB2API_OK proves that the extension, companion, and real provider page are connected.

API

Method	Path	Description
GET	/healthz	Daemon, extension connection, and ready worker status; no authentication
GET	/v1/models	Models available to the current accounts
POST	/v1/responses	Text, conversations, images, and function calls

Every endpoint except /healthz requires Authorization: Bearer <API_KEY>.

POST /v1/responses currently accepts:

Field	Behavior
model	Required; use an ID returned by /v1/models
input	A string or an array of input_text, input_image, and function_call_output items
instructions	Optional; applied to the emulated function-calling protocol
previous_response_id	Continue a web conversation on the same provider
stream	Set to true for text SSE
reasoning.effort	Strictly select an effort currently available in the page
tools	image_generation or function
tool_choice	auto, none, required, or a named function
parallel_tool_calls	Allow multiple calls in one response; defaults to true

Models

curl "$WEB2API_BASE_URL/models" \
  -H "Authorization: Bearer $WEB2API_KEY"

Stable default IDs are chatgpt/default, gemini/default, and grok/default. Dynamic account-visible models retain their provider prefix, such as gemini/3.1-pro or grok/4.5. Discover them through /v1/models instead of hard-coding them.

Text

curl "$WEB2API_BASE_URL/responses" \
  -H "Authorization: Bearer $WEB2API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemini/default",
    "input": "Explain why the sky is blue in two sentences."
  }'

Text is returned in output[].content[].text.

Streaming Text

curl -N "$WEB2API_BASE_URL/responses" \
  -H "Authorization: Bearer $WEB2API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "grok/default",
    "input": "Count from 1 to 5 slowly.",
    "stream": true
  }'

The response contains Responses API-style SSE events, including response.output_text.delta and response.completed.

Multi-turn Conversations

Pass a response id back as previous_response_id. The next request must use the same provider.

curl "$WEB2API_BASE_URL/responses" \
  -H "Authorization: Bearer $WEB2API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "chatgpt/default",
    "previous_response_id": "resp_chatgpt_REPLACE_ME",
    "input": "Now summarize your previous answer in one sentence."
  }'

Reasoning Effort

{
  "model": "grok/default",
  "input": "Solve this step by step: ...",
  "reasoning": {
    "effort": "expert"
  }
}

The effort must be a value currently reported for that provider in the popup. Web2API never silently falls back to another effort.

Image Understanding

input_image.image_url accepts HTTP(S) URLs and base64 data URLs. Each image may be up to 16 MiB, and multiple images are accepted in order. file_id is not supported.

curl "$WEB2API_BASE_URL/responses" \
  -H "Authorization: Bearer $WEB2API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemini/default",
    "input": [
      {"type": "input_text", "text": "Describe this image precisely."},
      {"type": "input_image", "image_url": "https://example.com/input.png"}
    ]
  }'

Convert a local image to a data URL with:

printf 'data:image/png;base64,'
base64 < input.png | tr -d '\n'

Image Generation

Declare the image_generation tool and use a non-streaming request. ChatGPT and Gemini use their respective image UI. Grok automatically opens Imagine.

curl "$WEB2API_BASE_URL/responses" \
  -H "Authorization: Bearer $WEB2API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "grok/default",
    "input": "A product photo of a transparent mechanical keyboard on a white desk, soft daylight",
    "tools": [{"type": "image_generation"}]
  }' > response.json

The final image is raw base64 in the image_generation_call.result output item:

{
  "type": "image_generation_call",
  "status": "completed",
  "result": "iVBORw0KGgo..."
}

Extract it with Node.js 22+:

node -e 'const fs = require("node:fs"); const body = JSON.parse(fs.readFileSync("response.json", "utf8")); const image = body.output.find((item) => item.type === "image_generation_call"); fs.writeFileSync("generated-image", Buffer.from(image.result, "base64"));'

Run file generated-image to identify the actual format, then add a .png or .jpg extension.

Generation features and quotas depend on the current provider account. If the provider page shows an upgrade, quota, or regional restriction, the API returns a provider error and does not bypass it.

Image Editing

Image editing combines input_image with image_generation. The input is uploaded to the real provider page and the edited image is returned:

curl "$WEB2API_BASE_URL/responses" \
  -H "Authorization: Bearer $WEB2API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "chatgpt/default",
    "input": [
      {"type": "input_text", "text": "Remove the background and keep the object unchanged."},
      {"type": "input_image", "image_url": "https://example.com/product.png"}
    ],
    "tools": [{"type": "image_generation"}]
  }'

To keep editing the generated image, include the previous response id as previous_response_id in the next request.

Function Calling

Web2API accepts Responses API function tools:

curl "$WEB2API_BASE_URL/responses" \
  -H "Authorization: Bearer $WEB2API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "chatgpt/default",
    "input": "What is the weather in Paris?",
    "tools": [{
      "type": "function",
      "name": "get_weather",
      "description": "Get current weather",
      "parameters": {
        "type": "object",
        "properties": {"city": {"type": "string"}},
        "required": ["city"],
        "additionalProperties": false
      },
      "strict": true
    }],
    "tool_choice": "auto"
  }'

Execute the returned function_call, then send its result with the original response id and call_id:

{
  "model": "chatgpt/default",
  "previous_response_id": "resp_chatgpt_REPLACE_ME",
  "input": [{
    "type": "function_call_output",
    "call_id": "call_REPLACE_ME",
    "output": "{\"temperature\":21,\"unit\":\"C\"}"
  }]
}

tool_choice: "auto" | "none" | "required", named functions, and parallel_tool_calls are supported. Function calling is emulated through a strict prompt protocol rather than native provider functions. strict: true cannot provide the same guarantee as OpenAI Structured Outputs. Tool-enabled streaming is buffered until the page response completes.

CLI

glidea-web2api prepare-extension
glidea-web2api doctor --extension-id <extension-id>
glidea-web2api start
glidea-web2api uninstall --extension-id <extension-id>

The extension popup controls the companion during normal use. start is only a manual fallback.

Development

Only source development requires a build:

pnpm install
pnpm build:daemon
pnpm build:extension
pnpm test
pnpm typecheck
pnpm test:e2e:extension

Load src/extension/.output/chrome-mv3, then run:

node bin/glidea-web2api.mjs install --extension-id <extension-id>

Real-page smoke tests use dedicated Chrome profiles under ~/.web2api/ and never copy cookies. See package.json for the commands.

Documentation

• Technical design
• Implementation tasks
• Local acceptance

Web2API is an unofficial local tool and is not affiliated with OpenAI, Google, or xAI. Use it in accordance with each provider's terms and account restrictions.