web-model

Turn real browser AI chat tabs into a local OpenAI-compatible API.

中文说明见 README.zh.md.

web-model keeps the real website, the real logged-in browser session, and the real page behavior.
The local server exposes registered tabs through /v1/models and /v1/chat/completions.

Highlights

• OpenAI-compatible GET /v1/models and POST /v1/chat/completions
• Real browser tabs, not private reverse-engineered APIs
• Chrome / Edge MV3 extension
• Streaming support
• Wildcard routing:
  • model: "*" picks any non-busy tab
  • model: "kimi*" picks any non-busy Kimi tab
• Error penalty:
  • a tab that returns a chat error gets a 1 minute penalty window
  • wildcard routing avoids penalized tabs when cleaner candidates exist
• Built-in TUI monitor with busy / usage / penalty countdown

Supported Sites

• ChatGPT
• Qwen
• Gemini
• Kimi
• Yuanbao

Architecture

your app / script
        |
        v
  web-model server
        |
        v
browser extension
        |
        v
 real browser tab
        |
        v
 provider website

Quick Start

1. Download and extract one package

tar -xzf macos-amd64-all.tgz
cd macos-amd64

Available packages:

• linux-amd64-all.tgz
• macos-amd64-all.tgz
• windows-amd64-all.tgz

2. Start the server

./web-model

On Windows:

.\web-model.exe

Default addresses:

• HTTP: http://127.0.0.1:18080
• WebSocket: ws://127.0.0.1:18080/ws

Help:

./web-model -h
./web-model -cn

3. Load the extension

Load the bundled extension/ directory as an unpacked extension in Chrome or Edge.

4. Prepare the provider page

The target page must already work by hand.

That includes:

• login
• auth
• onboarding
• consent dialogs
• model availability
• any site-specific page state

web-model only sends the prompt and parses the answer. It does not solve captcha, paywalls, product-side limits, or business rules.

Keep the page sidebar visible. If the left sidebar is collapsed, start new chat may fail.

5. Enable the tab

In the extension popup:

• Server URL = ws://127.0.0.1:18080/ws
• enable the current tab
• turn on debug logs only when needed

Badge meaning:

• Green ON: registered
• Yellow ...: connecting / waiting
• Red !: error
• Blank: disabled / inactive

6. Call the API

curl http://127.0.0.1:18080/v1/models

curl http://127.0.0.1:18080/v1/chat/completions \
  -H 'content-type: application/json' \
  -d '{
    "model": "kimi-tab-2123080689",
    "messages": [
      { "role": "user", "content": "hi" }
    ]
  }'

Model Routing

model supports three forms:

• Exact key
  • Example: kimi-tab-2123080689
• *
  • Randomly picks one non-busy registered tab
• xx*
  • Randomly picks one non-busy registered tab whose type is xx
  • Example: kimi*, qwen*, chatgpt*

Penalty behavior:

• if a tab returns a chat error, it gets a 1 minute penalty window
• wildcard routing avoids penalized tabs when other healthy candidates exist
• if only penalized candidates remain, they can still be used

The TUI shows this as the Penalty column.

Binaries

Release artifacts include three platform packages:

• linux-amd64-all.tgz
• macos-amd64-all.tgz
• windows-amd64-all.tgz

Each package contains:

• web-model or web-model.exe
• extension/

web-model also includes a built-in quick subcommand that calls the local API directly.

Examples:

./web-model quick kimi "hi"
./web-model quick --stream kimi "hi"
./web-model quick --stream '*' "hi"
./web-model quick --stream 'qwen*' "hi"

On Windows:

.\web-model.exe quick kimi "hi"
.\web-model.exe quick --stream '*' "hi"

Important:

• Quote * and xx* in your shell, or the shell may expand them into filenames.
• For wildcard requests, the server returns the actual concrete model key in the response.

API

• GET /healthz
• GET /providers
• GET /v1/models
• POST /v1/chat/completions

Streaming works with the standard OpenAI-style SSE response from /v1/chat/completions.

Supported request fields for POST /v1/chat/completions:

• model
• messages
• stream
• temperature
• top_p
• max_tokens
• stop
• user

Supported message fields:

• role
• content
• name

Supported response fields:

• id
• object
• created
• model
• choices
• usage

This is a useful OpenAI-compatible subset, not the full OpenAI API surface.

Notes

• Registration is per browser tab.
• Refreshing a page requires the extension bridge to reconnect.
• The extension maintains a heartbeat for registered tabs so idle browser pages are less likely to silently disappear.
• Streaming quality depends on the provider site. Final non-streamed responses are usually cleaner.

License

MIT