SubscriptionBridge — OpenCode Integration

Overview

SubscriptionBridge exposes an OpenAI-compatible API at /v1 so that OpenCode and other OpenAI-compatible clients can use browser-based LLMs (Gemini, ChatGPT) through SubscriptionBridge.

Architecture when connected to OpenCode:

OpenCode is the coding agent — it handles project context, tools, file edits, shell commands, and code patches.
SubscriptionBridge is the model/backend bridge — it receives chat completion requests from OpenCode and forwards them to the browser-based provider.

OpenCode Configuration

Add the following to your OpenCode configuration (e.g., ~/.config/opencode/opencode.json):

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "subscription-bridge": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "SubscriptionBridge Local",
      "options": {
        "baseURL": "http://127.0.0.1:8787/v1",
        "apiKey": "dummy"
      },
      "models": {
        "subscription-bridge-fake": {
          "name": "SubscriptionBridge Fake",
          "limit": { "context": 32000, "output": 8192 }
        },

        "subscription-bridge-gemini-fast": {
          "name": "Gemini 3 Flash",
          "limit": { "context": 1000000, "output": 8192 },
          "modalities": { "input": ["text", "image"], "output": ["text"] }
        },
        "subscription-bridge-gemini-thinking": {
          "name": "Gemini 3 Deep Think",
          "limit": { "context": 192000, "output": 65536 },
          "modalities": { "input": ["text", "image"], "output": ["text"] }
        },
        "subscription-bridge-gemini-pro": {
          "name": "Gemini 3.1 Pro",
          "limit": { "context": 1000000, "output": 65536 },
          "modalities": { "input": ["text", "image"], "output": ["text"] }
        },

        "subscription-bridge-chatgpt": {
          "name": "ChatGPT",
          "limit": { "context": 128000, "output": 16384 },
          "modalities": { "input": ["text", "image"], "output": ["text"] }
        }
      }
    }
  }
}

Compaction Config

Recommended to include with OpenCode:

{
  "compaction": {
    "auto": true,
    "prune": true,
    "reserved": 100000
  }
}

OpenCode manages context and session history. Each request to SubscriptionBridge uses a fresh chat or guaranteed reset. OpenCode is the source of truth.

Subagent Model Mapping

OpenCode can use different models for different agent types. Recommended configuration:

Agent Role	Gemini Option	ChatGPT Option
`build`	`subscription-bridge-gemini-pro`	`subscription-bridge-chatgpt`
`plan`	`subscription-bridge-gemini-thinking`	`subscription-bridge-chatgpt`
`explore`	`subscription-bridge-gemini-fast`	`subscription-bridge-chatgpt`
`review`	`subscription-bridge-gemini-pro`	`subscription-bridge-chatgpt`

SubscriptionBridge does not implement subagents. OpenCode manages agents.

Setup Steps

Option A: Gemini

# 1. Start Chrome with remote debugging
google-chrome \
  --remote-debugging-port=9333 \
  --user-data-dir="$HOME/.subscription-bridge/chrome-profile" \
  --no-first-run \
  --no-default-browser-check

# 2. Navigate to https://gemini.google.com/app and log in manually

# 3. Start the bridge server
bridge server --provider gemini

# 4. Verify models
curl http://127.0.0.1:8787/v1/models

# 5. Select a Gemini model in OpenCode

Option B: ChatGPT

# 1. Start Chrome with remote debugging
google-chrome \
  --remote-debugging-port=9333 \
  --user-data-dir="$HOME/.subscription-bridge/chrome-profile" \
  --no-first-run \
  --no-default-browser-check

# 2. Navigate to https://chatgpt.com and log in manually

# 3. Start the bridge server
bridge server --provider chatgpt

# 4. Verify models
curl http://127.0.0.1:8787/v1/models

# 5. Select the ChatGPT model in OpenCode

Option C: Both

# Start server with both providers
bridge server --provider both

# Or let the server prompt you
bridge server
# ? Which provider do you want to use?
#   1) Gemini
#   2) ChatGPT
#   3) Both

Model Reference

Model ID	Backend	Context	Output	Browser Required
`subscription-bridge-fake`	FakeProviderAdapter	32K	8K	No
`subscription-bridge-gemini-fast`	Gemini 3 Flash	1M	8K	Yes
`subscription-bridge-gemini-thinking`	Gemini 3 Deep Think	192K	64K	Yes
`subscription-bridge-gemini-pro`	Gemini 3.1 Pro	1M	64K	Yes
`subscription-bridge-chatgpt`	ChatGPT (GPT-4o)	128K	16K	Yes

Tool-Call Compatibility

SubscriptionBridge supports OpenAI tool calls in the following way:

Tools are accepted in /v1/chat/completions requests
Tools are converted into prompt instructions — the model is told to respond with a JSON tool_calls object when a tool is needed
The model's JSON output is parsed and converted to OpenAI-compatible tool_calls in the response
Tool results from OpenCode (role: "tool") are converted into readable transcript entries in the prompt
SubscriptionBridge does not execute tools when used via OpenCode — OpenCode handles execution

When the model decides to call a tool, the response contains:

{
  "choices": [{
    "finish_reason": "tool_calls",
    "message": {
      "role": "assistant",
      "content": null,
      "tool_calls": [{
        "id": "call_1",
        "type": "function",
        "function": {
          "name": "tool_name",
          "arguments": "{\"arg\": \"value\"}"
        }
      }]
    }
  }]
}

Known Limitations

Tool call reliability: Browser-based models may not reliably produce structured tool call JSON. Results may vary compared to API-based models.
Streaming is simulated: True token-by-token streaming from the browser provider is not available. Response text is chunked into SSE events.
Token counts are approximate: Uses character count / 4, not a real tokenizer.
Context limit is approximate: Based on estimated tokens. If exceeded, the API returns context_length_exceeded — OpenCode compaction handles this.
Browser latency: Responses are slower than API-based models.
No multimodal: Image inputs in content arrays are accepted but ignored.
API key: Any value works (e.g., "dummy"). The server does not validate.

FilesExpand file tree

opencode.md

Latest commit

History