NOAPI?

Use ChatGPT, Claude, Gemini, DeepSeek and friends inside OpenCode, with no API keys.

Your browser is already logged in. NOAPI? just lets OpenCode talk to it.

Quick start  •  How it works  •  Any project folder  •  Troubleshooting  •  Limitations

💡 Why this exists

API access costs money, and a lot of us already pay for a chat subscription that we barely squeeze the value out of. NOAPI? connects the two. OpenCode sends a prompt, a small browser extension types it into the real chat page you have open, waits for the answer, and hands it back.

It is not magic and it is not an official integration. It is a careful bit of plumbing that works well enough that you can run a real coding agent on top of it, including file edits and shell commands.

✨ What you get

🌐 Supported sites

	Site	Model id in OpenCode	Address
	ChatGPT	noapi/chatgpt	chatgpt.com
	Claude	noapi/claude	claude.ai
	Gemini	noapi/gemini	gemini.google.com
	DeepSeek	noapi/deepseek	chat.deepseek.com
	Perplexity	noapi/perplexity	perplexity.ai
	Grok	noapi/grok	grok.com
	Mistral	noapi/mistral	chat.mistral.ai

You need to be logged in, and the site has to be open in a tab. If you open several, each one can serve its own model.

🚀 Quick start

You need: Node.js 18 or newer, Chrome (or Edge, Brave, Opera, Arc) or Firefox, and OpenCode.

1️⃣ Load the extension

2️⃣ Start the bridge

npm start

Leave it running. You can peek at the dashboard any time at http://127.0.0.1:8787/.

3️⃣ Open an AI site

Open one of the supported sites and log in. You will see a small round ? button in the bottom right corner.

4️⃣ Use it in OpenCode

Run opencode inside this folder, type /models, search for NOAPI? and pick something like noapi/chatgpt. Then just work as usual.

💬 Try: "make a folder named test and put a hello.txt inside it". If everything is wired up, you will see the status lines scroll by and the files show up on disk.

📁 Use it from any project folder

The opencode.json in this repo only applies when you launch OpenCode from this exact folder. To use NOAPI? in all your projects, register it once in your global OpenCode config:

npm run install-global

That adds the noapi provider and the MCP server to ~/.config/opencode/opencode.json, using the absolute path of this folder. Your other settings are kept, and a backup is saved next to the file as opencode.json.bak.

Then, for each project:

# terminal 1, inside the NOAPI? folder
npm start

# terminal 2, inside your project
opencode

A few things worth knowing:

• Start the bridge before OpenCode. If OpenCode has to start it for you and you close that session, other sessions lose the model.
• If you move the NOAPI? folder, run npm run install-global again.
• To undo it, run npm run install-global -- --remove.

🧩 How it works

flowchart LR
    A["OpenCode"] -- "OpenAI-style request<br>POST /v1/chat/completions" --> B["Local bridge<br>127.0.0.1:8787"]
    A -. "MCP over stdio" .-> B
    B -- "WebSocket" --> C["Browser extension"]
    C -- "types, sends, reads" --> D["ChatGPT / Claude / Gemini<br>DeepSeek / ..."]
    D -- "reply" --> C
    C --> B
    B -- "answer or tool call" --> A

In plain words:

1. OpenCode thinks it is talking to an OpenAI compatible server. That server is the bridge, a tiny Node program on your machine.
2. The bridge forwards the prompt over a WebSocket to the extension.
3. The extension finds the chat box on the open tab, types the prompt, presses send, and watches the page until the answer is complete.
4. The answer travels back the same way.

🧰 How tools work on a website that has no tools

Chat websites do not accept tool definitions, so the bridge improvises:

1. On the first prompt of a conversation it adds a short briefing that lists the available tools and a simple format for calling them.
2. When the model wants to act, it writes a fenced block starting with TOOL_CALL, followed by a small JSON object.
3. The bridge cuts that block out of the reply and hands OpenCode a real tool_calls response.
4. OpenCode runs the tool and sends the result back, which the bridge pastes into the chat as the next message.

Because the website already remembers the conversation, only the new message is sent each turn, not the whole history. Very long tool output is trimmed in the middle to keep prompts sane.

🛡️ Small details that make it nicer

• 🚦 Requests to the same site are queued, so two prompts never collide in one chat box.
• ⏹️ Press Esc in OpenCode and the extension also hits the stop button on the website.
• 🏷️ OpenCode's session title requests are answered locally, so they do not clutter your real chat.
• 🔁 If a site swallows a send, or wipes the text you just typed, the extension notices and tries again.

⚙️ Configuration

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "noapi": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "NOAPI? (browser bridge)",
      "options": {
        "baseURL": "http://127.0.0.1:8787/v1",
        "apiKey": "noapi-local"
      },
      "models": {
        "chatgpt":    { "name": "ChatGPT via browser",    "limit": { "context": 128000,  "output": 8192 } },
        "claude":     { "name": "Claude via browser",     "limit": { "context": 200000,  "output": 8192 } },
        "gemini":     { "name": "Gemini via browser",     "limit": { "context": 1000000, "output": 8192 } },
        "deepseek":   { "name": "DeepSeek via browser",   "limit": { "context": 128000,  "output": 8192 } },
        "perplexity": { "name": "Perplexity via browser", "limit": { "context": 100000,  "output": 8192 } },
        "grok":       { "name": "Grok via browser",       "limit": { "context": 128000,  "output": 8192 } },
        "mistral":    { "name": "Mistral via browser",    "limit": { "context": 128000,  "output": 8192 } }
      }
    }
  },
  "mcp": {
    "noapi": {
      "type": "local",
      "command": ["node", "/absolute/path/to/NOAPI/bridge/index.js", "--mcp"],
      "enabled": true,
      "timeout": 30000
    }
  }
}

node bridge/index.js [--mcp] [--port=8787] [--host=127.0.0.1]

🔧 Troubleshooting

🪪 Tip: keep the AI tab visible when you can. Browsers slow down background tabs, which can make long responses laggy.

🧪 Testing without a browser

You can check the whole pipeline with a fake browser tab. No real site needed.

npm test

This starts a bridge, connects a fake extension, and runs plain requests, streaming, MCP in both modes, and a full tool call round trip (model asks for bash, gets the result, answers).

🗂️ Project layout

NOAPI/
├── bridge/                 Local server (pure Node.js)
│   ├── index.js            Entry point, WebSocket + HTTP + MCP wiring
│   ├── http.js             OpenAI compatible endpoints and dashboard
│   ├── state.js            Request queue, routing, progress, cancel
│   ├── toolproto.js        TOOL_CALL protocol: prompt building and parsing
│   ├── mcp.js              MCP server over stdio (with proxy mode)
│   ├── ws.js / wsclient.js Minimal WebSocket server and client
│   └── sites.js            Supported sites
├── extension/
│   ├── chrome/             Manifest V3 extension
│   └── firefox/            Same code, Firefox manifest
├── tools/
│   ├── install-global.js   Register NOAPI? in your global OpenCode config
│   ├── fake-extension.js   Fake browser tab for tests
│   ├── test-all.js         End to end tests
│   └── test-tools.js       Tool call round trip test
└── opencode.json           Project level OpenCode config

🤝 Adding or fixing a site

Sites change their HTML from time to time, so selectors rot. The good news is that each site is only a small block of config.

1. In extension/chrome/content.js, find the SITES object. Each entry lists CSS selectors for the input, the send button, the stop button and the reply messages.
2. Open the site, right click the element, choose Inspect, and update the selector.
3. Copy the same file to extension/firefox/content.js so both browsers match.
4. For a brand new site, also add it to bridge/sites.js, to opencode.json, and to the matches and host_permissions in both manifests.
5. Run npm test, reload the extension, and try it on the real page.

Pull requests with selector fixes are very welcome. 🙏

⚠️ Honest limitations

I would rather tell you now than have you find out later.

• It depends on page layouts. There is no official interface here. When a site redesigns its chat page, that site can break until the selectors are updated.
• Speed. You are driving a real web page, so replies take as long as the website takes, plus a little overhead.
• Tool quality varies by model. Tool use works through a text format the model has to follow. Strong models do it reliably, weaker ones occasionally slip.
• One conversation per site tab. The website keeps the history, so a very long session may hit that site's own limits.
• Free plans have limits too. The websites still apply their own message caps and rules.
• Terms of service. Automating a website may go against that site's terms. This project is meant for personal use with your own account. You are responsible for how you use it, and it is not affiliated with OpenAI, Anthropic, Google, DeepSeek, Perplexity, xAI, Mistral or OpenCode.

🔐 Privacy and security

• 🏠 The bridge listens on 127.0.0.1 only. Nothing is exposed to your network.
• 🚫 No analytics, no tracking, no external servers.
• 🍪 The extension never reads or stores your passwords or cookies. It only types into and reads from the chat page you already opened.
• 👀 It is plain JavaScript with no build step, so what you read is what runs.

📄 License

NOAPI? is released under the MIT License. Use it, change it, share it. Just keep the license notice.