NanoProxy

NanoProxy is a bridge layer for NanoGPT when native tool calling is unreliable.

It sits between the client and NanoGPT, replaces fragile native tool-calling with a stricter bridge protocol, and converts the result back into normal OpenAI-style tool_calls for the client. That lets tools like OpenCode keep working normally even when NanoGPT would otherwise stop early, leak raw tool text, or return malformed tool output.

Which mode should I use?

Use the plugin if you use OpenCode

This is the easiest setup for OpenCode.

• No local proxy server to keep running
• No custom provider setup
• Keep using the normal built-in NanoGPT provider

Use the standalone server for other tools

Use this if:

• you are not using OpenCode
• your client supports an OpenAI-compatible base URL
• you prefer a separate local proxy process

OpenCode Plugin Setup

The plugin intercepts NanoGPT API requests inside OpenCode and applies the NanoProxy bridge automatically.

Example config:

{
  "plugin": [
    "file:///path/to/NanoProxy/src/plugin.mjs"
  ]
}

Notes:

• Use a real absolute file path.
• On Windows, a valid example looks like:
  • file:///C:/Users/you/path/to/NanoProxy/src/plugin.mjs
• After editing the config, restart OpenCode.

Plugin debug logging

Enable debug logging for one OpenCode run:

NANOPROXY_DEBUG=1 opencode

Optional:

• set NANOPROXY_LOG=/path/to/file to change the single event log file location
• set NANOPROXY_LOG_DIR=/path/to/folder to change where detailed per-request debug files are written

On Windows, you can also use the same persistent toggle used by the standalone server:

./toggle-debug.ps1

That writes a .debug-logging flag file in the repo. When that flag is present, both:

• the OpenCode plugin
• the standalone server

will enable NanoProxy debug logging until you toggle it off again.

When debug mode is enabled, NanoProxy also writes:

• a single event log file
• per-request *-request.json
• raw streamed *-stream.sse
• parsed *-response.json

By default these go into:

• event log: your system temp folder as nanoproxy-plugin.log
• detailed logs: your system temp folder under nanoproxy-plugin-logs

Standalone Server Setup

Run the server:

node server.js

Then point your coding tool to:

http://127.0.0.1:8787

Keep using your normal NanoGPT API key in that tool.

If your tool supports a custom OpenAI-compatible provider or baseURL, use http://127.0.0.1:8787 there.

Optional overrides:

UPSTREAM_BASE_URL=https://nano-gpt.com/api/v1
PROXY_HOST=127.0.0.1
PROXY_PORT=8787
node server.js

Server debug logging

Off by default.

Enable for one run:

NANO_PROXY_DEBUG=1 node server.js

Or toggle persistently on Windows:

./toggle-debug.ps1

That same toggle also enables plugin debug logging.

Server logs are written to Logs/.

Health check

curl http://127.0.0.1:8787/health

Docker

If you want to run the standalone server in Docker instead of running Node directly:

docker build -t nano-proxy .
docker run --rm -p 8787:8787 nano-proxy

Or with Compose:

docker compose up --build

This still exposes the proxy at:

http://127.0.0.1:8787

What NanoProxy actually does

For tool-enabled requests:

1. It removes the normal native tool-calling structure before sending the request upstream.
2. It tells the model to use a stricter text-based tool format instead.
3. It watches the model output.
4. It converts that output back into normal OpenAI-style tool_calls.

So your client still sees normal tool calls, but NanoGPT does not have to rely on its native tool-calling behavior.

Bridge format

Tool reply:

[[OPENCODE_TOOL]]
[[CALL]]
{"name": "read", "arguments": {"filePath": "src/app.js"}}
[[/CALL]]
[[/OPENCODE_TOOL]]

Multiple independent tool calls in one turn:

[[OPENCODE_TOOL]]
[[CALL]]
{"name": "read", "arguments": {"filePath": "src/app.js"}}
[[/CALL]]
[[CALL]]
{"name": "read", "arguments": {"filePath": "src/styles.css"}}
[[/CALL]]
[[/OPENCODE_TOOL]]

Final answer:

[[OPENCODE_FINAL]]
Your answer here.
[[/OPENCODE_FINAL]]

Notes

• Requests without tools are forwarded unchanged.
• Reasoning streams live.
• Tool and final content are buffered until NanoProxy can classify them safely.
• By default, NanoProxy allows up to 5 tool calls in a single assistant turn for models that behave well with batching.
• Some models may still behave better with one tool call per turn.

Troubleshooting

OpenCode says the config file is invalid JSON

• Make sure the config file is valid JSON or JSONC.
• On Windows, avoid saving it with weird encoding or a BOM if OpenCode rejects it.

Plugin seems to do nothing

• Check that the plugin path is correct.
• Make sure it is a file:///... URL, not a normal path string.
• Restart OpenCode after editing the config.
• If needed, enable NANOPROXY_DEBUG=1 and check the plugin log output.

Tool text leaks into the chat

• This usually means the model drifted into a malformed tool format.
• Try again once first.
• If it keeps happening, enable debug logs and inspect what the model actually emitted.

Standalone server does not respond

• Make sure node server.js is still running.
• Check http://127.0.0.1:8787/health.

Project Structure

NanoProxy/
|-- server.js
|-- src/
|   |-- core.js
|   `-- plugin.mjs
|-- selftest.js
|-- README.md
|-- package.json
|-- Dockerfile
`-- docker-compose.yml

Verification

node --check server.js
node selftest.js

License

MIT