Developer tooling for agent systems — context enforcement, test harnesses, protocol gateways, audit trails, and local workflows you can actually inspect.
Models are useful; the infrastructure around them is hard to test, constrain, and recover from when things fail. That's the gap I work in.
Self-hosted context enforcement for LLM APIs, MCP, vLLM, Ollama, and agent traffic. Local proxy or stdio MCP wrapper; tokenize/redact payloads, audit without raw secrets or PII.
Lightweight Go gateway for A2A traffic — rate limiting, auth modes, Agent Card checks, audit logs, metrics, hot reload. Start here, grow into agentgateway.
Windows desktop terminal multiplexer for running AI agent sessions, shells, and browser-assisted workflows side by side. Built around workspaces, split panes, session recovery, and agent status visibility.
Local artifact exchange for LLM workflows over HTTP, CLI, and MCP. Agents can publish, read, update, and continue from the same artifact without copying context through chat.
Local harness around OpenAI Codex CLI. Durable run ledgers, verification-gated completion, bounded repair loops. "Done" means the check passed.
CLI for MCP server developers: inspect servers, generate YAML test specs, run assertions, CI-friendly output.
- VS Code extension — run specs and see failures in the editor
- MCP server adapter — agents inspect, generate, and run tests as tools