I build both conventional LLM integrations and heavier agent systems.
On agent projects, my usual pattern is a small self-contained MCP server around the agent's real tools and context, with task skills telling the agent which MCP tools to use. My usual runtime layer is Codex SDK / OpenCode SDK or the same class of wrapper; Codex App Server comes in when the pipeline is heavy enough to justify it.
Typical scope can include:
- Go or TypeScript MCP servers around your product tools and APIs
- Agent workflows using Codex SDK / OpenCode SDK-style runtimes
- Structured model outputs and validation where the task needs them
- Memory or retrieval where the product actually needs them
- Run history, prompt/skill versions, cost tracking and eval/debug tooling
- Integration with the existing backend and deployment setup
My public Fixer MCP project is an experimental control plane I use daily across 32+ projects. SoulWi is the commercial conventional-LLM example on my profile.
