Skip to main content

Programmatic Adapter Opportunities Research

This document summarizes research into which existing subprocess-based adapters would benefit from programmatic SDK or remote API variants.

Current State​

Existing Adapter Types​

  • Subprocess adapters (11): cursor, copilot, gemini, qwen, hermes, omp, openclaw, claude, codex, pi, opencode
  • Programmatic adapters (3): claude-agent-sdk, codex-sdk, pi-sdk
  • Remote adapters (3): opencode-http, codex-websocket, agent-mux-remote

Research Results​

High Priority Candidates​

1. Google Gemini SDK Adapter​

  • Current: gemini CLI command
  • Opportunity: Official Google AI SDKs (Python, Node.js, Go, Java)
  • Benefits:
    • Direct SDK integration eliminates CLI subprocess overhead
    • Better streaming support with WebSocket via Gemini Live API
    • Native tool calling without CLI parsing
    • Production-ready officially supported libraries
  • Recommendation: Implement gemini-sdk-adapter extending BaseProgrammaticAdapter

2. Qwen/DashScope SDK Adapter​

  • Current: qwen CLI command
  • Opportunity: Official Alibaba DashScope SDKs + OpenAI-compatible API
  • Benefits:
    • Already uses OpenAI-compatible API underneath CLI
    • Official Python/Node.js SDKs available
    • Async/streaming support built-in
    • Significant performance improvement over CLI
  • Recommendation: Implement qwen-sdk-adapter extending BaseProgrammaticAdapter

3. GitHub Copilot SDK Adapter​

  • Current: gh copilot CLI extension
  • Opportunity: New GitHub Copilot SDK (public preview as of April 2026)
  • Benefits:
    • Same engine as CLI but with programmatic control
    • Multi-language SDKs (Python, TypeScript, Go, .NET, Java)
    • OpenTelemetry support for distributed tracing
    • JSON-RPC communication with Copilot server
  • Recommendation: Implement copilot-sdk-adapter extending BaseProgrammaticAdapter

Medium Priority Candidates​

4. Cursor API Adapter​

  • Current: cursor-agent CLI command
  • Opportunity: Community reverse-engineered APIs (HTTP/2 + streaming)
  • Status: No official SDK, but APIs exist
  • Recommendation: Monitor for official API release before implementing

5. Hermes RPC Adapter​

  • Current: hermes CLI command
  • Opportunity: Proposed RPC mode + existing REST endpoints
  • Status: RPC mode not yet implemented
  • Recommendation: Consider when RPC mode becomes available

6. OpenClaw SDK Adapter​

  • Current: openclaw CLI command
  • Opportunity: Python SDK (openclaw-py) + REST API
  • Benefits: Open-source with plugin system
  • Recommendation: Medium priority for open-source integration scenarios

7. OMP/MCP​

  • Analysis: MCP is a protocol standard, not an agent platform
  • Current Support: Already integrated in existing adapters via MCP plugin support
  • Recommendation: No separate adapter needed

Implementation Recommendations​

Phase 1: High-Impact SDK Adapters​

  1. gemini-sdk-adapter - Google AI SDK integration
  2. qwen-sdk-adapter - DashScope SDK integration
  3. copilot-sdk-adapter - GitHub Copilot SDK integration

Phase 2: API/Remote Adapters​

  1. cursor-api-adapter - When official API becomes available
  2. hermes-rpc-adapter - When RPC mode is implemented

Phase 3: Specialty Adapters​

  1. openclaw-sdk-adapter - For open-source integration needs

Technical Approach​

Programmatic Adapters​

  • Extend BaseProgrammaticAdapter
  • Use official SDKs for direct integration
  • Implement streaming via SDK native capabilities
  • Mock implementations for testing

Remote Adapters​

  • Extend BaseRemoteAdapter
  • HTTP/WebSocket connections to APIs
  • Server lifecycle management
  • Real-time bidirectional communication

Next Steps​

  1. Implement high-priority SDK adapters (gemini-sdk, qwen-sdk, copilot-sdk)
  2. Update adapter validation to support non-subprocess adapter types
  3. Create comprehensive tests for new adapter variants
  4. Update documentation with new adapter capabilities
  5. Monitor Cursor and Hermes for official API releases

Sources​