A modular Go runtime for the Model Context Protocol (MCP) — designed for intelligent agent coordination, model selection, prompt optimization, and lifecycle-aware execution.
This SDK helps you build systems that:
- Coordinate multiple models (e.g., GPT-4o, Claude, Gemini)
- Dynamically generate and execute agents
- Route tasks based on trust, capability, or intent
- Manage prompt formatting and lifecycle control
📖 Read the full article on LinkedIn →
Modern AI infrastructure is agentic: prompts now flow through models, memory, tools, and reasoning layers. MCP offers a universal protocol for managing that flow.
This Go implementation provides:
- Concurrency-safe, production-ready execution
- Symbolically annotated agent flows
- Pluggable model and tool support
- Session lifecycle and halting controls
SessionCoordinatorfor request routing, lifecycle hooksModelRouterfor dynamic model selectionAgentGeneratorfor building runtime agents on demandPromptOptimizerStepto transform user input into optimized promptsHaltingPolicyto bound executionDescriptorHash()to lock and validate execution configuration
go get github.com/modelcontextprotocol/mcp-agent-runtime-go
/core/ # Protocol definitions, session types
/model/ # ModelSpec, Registry, Router logic
/agentgen/ # Dynamic agent generation
/promptopt/ # Prompt optimization steps and templates
/ext/ # Optional extensions (e.g., memory, auth)
/transport/ # HTTP/SSE/WebSocket integrations
roadmap.md # Strategic plan for contributors
README.md # This file
See roadmap.md for full scoring and implementation priority.
| Core Module | Priority | Description |
|---|---|---|
| SessionLifecycle | 10 | Session entry points and lifecycle events |
| ModelRouter | 10 | Multi-model dispatch logic |
| AgentGenerator | 10 | Runtime agent composition and execution |
| HaltingPolicy | 9 | Safe runtime control |
| PromptOptimizerStep | 9 | Transform intent into optimized prompts |
| DescriptorHash() | 10 | Lock agent and routing structure |
This repo is open for discussion and design iteration.
- Start with GitHub Discussion #364
- See roadmap.md for priorities
- Use structured PRs based on core module folders
MIT License — open use encouraged with attribution.