OpenCode Ollama Tool-Call Proxy Plugin

An OpenCode plugin that does two things:

1. Tool-call conversion: intercepts chat-completion responses from a local Ollama server and rewrites raw tool-call markup into standard OpenAI tool_calls chunks. This lets models that do not expose native Ollama tool support still work as agentic coding models inside OpenCode.
2. Line-ending normalization: normalizes LF/CRLF line endings on write, edit, and multiedit tool arguments, and on files after edits, using .editorconfig or the OPENCODE_LINE_ENDINGS environment variable. Platform default: CRLF on Windows, LF on Linux/macOS.

Supported model families

| Family | Example Ollama tags | Raw markup handled | |--------|---------------------|--------------------| | Qwen 2.5 / Qwen 2.5-Coder | qwen2.5-coder, qwen2.5 | <tools>, <tool_call>, raw JSON object/array | | Qwen 3.5 / Qwen 3-Coder | qwen3.5, qwen3-coder | <tool_call>{...}</tool_call> and <function=...><parameter=...> XML | | Ornith 1.0 | ornith-9b, ornith-35b | same as Qwen 3.5 (routed to Qwen converter) | | Gemma 4 | gemma4 | <|tool_call>call:name{...}<tool_call|>, <start_function_call>...<end_function_call> | | Mistral / Ministral | mistral, mistral-nemo, ministral, ministral-3 | [TOOL_CALLS] JSON array and <name>[ARGS]{...} | | Llama 3.x / OpenCoder | llama3.1, llama3.2, llama3.3, opencoder | raw JSON object/array, <|python_tag>...<|eom_id|> | | Phi-4 / Phi-4 Tools | phi4-mini, phi4-tools | {"name":..., "parameters":...}, functools[...] | | DeepSeek R1 / V3 | deepseek-r1, deepseek-v3 | <｜tool▁call▁begin｜> tags, fenced JSON, raw JSON, strips <think> | | Yi-Coder | yi-coder | raw JSON / JSON array, code fences | | GLM-4 / ChatGLM | glm4, chatglm3 | [TOOL_REQUEST] {...} [END_TOOL_REQUEST], legacy <tool_call> XML | | Laguna XS 2.1 | laguna-xs-2.1 | [tool=name<Prop ... />], [tool=name<arg_key>k</arg_key><arg_value>v</arg_value>] |

Installation

From npm (recommended)

npm install -g opencode-ollama-toolcall-proxy

Then add the package name to your OpenCode config:

{
  "plugin": ["opencode-ollama-toolcall-proxy"]
}

Restart OpenCode. The plugin is loaded automatically.

From a local checkout

git clone https://github.com/YOUR_USER/opencode-ollama-toolcall-proxy.git
npm install
npm run build

Reference the built entry point with an absolute file:// URL:

{
  "plugin": ["file:///C:/absolute/path/to/opencode-ollama-toolcall-proxy/dist/index.js"]
}

Usage

1. Make sure Ollama is running.
2. Configure your model in ~/.config/opencode/opencode.json with "tools": true:

{
  "provider": {
    "ollama": {
      "npm": "@ai-sdk/openai-compatible",
      "options": { "baseURL": "http://localhost:11434/v1" },
      "models": {
        "qwen2.5-coder:14b": {
          "name": "Qwen2.5 Coder 14B",
          "tools": true,
          "thinking": false
        }
      }
    }
  }
}

3. Start OpenCode. When the model emits raw tool-call text, the plugin converts it to proper tool_calls before the AI SDK sees the response.

Line-ending configuration

The plugin normalizes line endings on file writes/edits in this priority order:

1. OPENCODE_LINE_ENDINGS environment variable (lf or crlf)
2. .editorconfig end_of_line setting (walks up the directory tree)
3. Platform default: crlf on Windows, lf on Linux/macOS

Example .editorconfig:

root = true

[*]
end_of_line = crlf

[*.sh]
end_of_line = lf

Development

# type check
npm run typecheck

# run tests
npm test

# build a single ESM bundle for publishing
npm run build

Publishing

1. Pick a unique npm package name and update name in package.json.
2. Log in to npm: npm login.
3. Publish: npm publish.

The prepublishOnly script automatically builds dist/index.js before the package is uploaded.

License

MIT