I build the infrastructure that lets AI agents write code without being trusted to grade their own work: MCP gateways, agent harnesses, and verification that comes from exit codes rather than a model's opinion.
Nine years of .NET and C# before that, mostly ERP integration delivered at customer sites. Now working in Rust and TypeScript, directed rather than hand-written, which is exactly why the verification layer matters to me.
-
mcp-proxy — fork of joshrotenberg/mcp-proxy adding lazy backend spawning with a persistent warm tool catalog (list requests answered with zero processes running), a shared backend pool, endpoint groups with namespace-isolated capability filtering, and MCP 2026-07-28 protocol support. Design decisions are written up in
docs/adr/. -
avalonia-ui-mcp-rs — five-crate Rust workspace published on crates.io (core, services, tools, resources, server), MIT, MSRV 1.85. 18+ MCP tools for AvaloniaUI development; a full Rust rewrite of decriptor/AvaloniaUI.MCP.
-
Rust support in mcp-migration-check (merged PR) — rule MCP010: reads
Cargo.tomland flags Rust SDKs that cannot serve MCP 2026-07-28, with remediation forrmcp,rust-mcp-sdk, and thetower-mcpprotocol-2026-07-28feature.
- microsoft/vscode#319592
— reported and diagnosed: Copilot Chat threw away valid generated text when a
custom endpoint returned
finish_reason: "length", treating a truncated completion as a hard failure instead of rendering the partial response. Fixed in VS Code.
Writing about why the agent that reports a verdict must never be the agent that computes it, and looking for remote work on MCP servers, agent harnesses, and AI infrastructure.