Add OpenAI-compatible /v1/tools REST endpoints alongside MCP

[BillJr99]

Summary

This PR adds OpenAI-compatible REST endpoints (GET /v1/tools and POST /v1/tools/{name}/invoke) to the frontend UI server, allowing tools to be listed and invoked without speaking the MCP protocol. The MCP endpoint and all existing functionality remain completely unchanged.

Key Changes

• New tool_registry.py module: A shared in-process tool store that both server.py and frontend/app.py can access. server.py populates it during tool registration; frontend/app.py reads from it to serve the OpenAI endpoints.

• OpenAI-compatible endpoints in frontend/app.py:

  • GET /v1/tools — returns all registered tools in OpenAI function-calling schema format
  • POST /v1/tools/{tool_name}/invoke — invokes a tool with {"arguments": {...}} and returns a normalized tool_result response
  • Result normalization: strings are wrapped in a text content block; dicts with a content key pass through; other dicts are JSON-serialized; lists are used directly
  • Error handling: tool not found → HTTP 404; handler exceptions → HTTP 200 with is_error: true so LLM callers see errors as tool results

• Updated server.py: Calls tool_registry.register() inside register_tool() for every enabled tool, populating the shared registry.

• Enhanced tests/test_mcp_client.sh:

  • Added protocol mode selection (MCP vs OpenAI) before tool listing
  • Conditional logic to use either JSON-RPC /mcp calls or REST /v1/tools calls based on selected mode
  • Both modes converge to the same tool selection and result display logic
  • File listing step (mcpproxy__listfiles / mcpproxy__getfile) works with both protocols

• New test file tests/test_openai_api.py: Comprehensive test suite for the OpenAI endpoints covering:

  • Tool listing (empty registry, single/multiple tools, schema validation)
  • Tool invocation (success cases, result normalization, error handling)
  • Handler context passing and exception safety

• New tests in tests/test_server.py: Unit tests for the tool_registry module (register, get, get_all, clear, overwrite behavior).

Implementation Details

• The shared registry uses a plain module-level dict since both server.py and frontend/app.py run in the same Python process.
• The OpenAI endpoints are a pure addition — the MCP protocol path is completely unaffected.
• Handler invocation passes ctx=None to the dynamic_tool callable, which is safe for tools that don't use MCP context features.
• The test script now supports both MODE=1 (MCP) and MODE=2 (OpenAI) environment variable overrides to skip the interactive mode menu.

https://claude.ai/code/session_01V9SNVHv8uhS8pejXW9Dt7m