I build infrastructure and tools for AI agents.
My work focuses on reducing context overhead, preserving exact recovery, routing web search, and making agent workflows easier to inspect and reproduce.
A local stdio MCP proxy that reduces Tool catalogs and eligible oversized results before they reach the agent. It stores immutable result capsules so exact recovery does not execute the upstream Tool again.
The frozen bilingual fixture shows 76.18% fewer normal-path protocol tokens. A separate synthetic stress test reached 99.71%. These are fixture-specific results, not universal savings claims.
A free-first, self-hosted web-search router. It combines zero-key sources with explicit policies for optional paid providers and returns compact, bounded evidence.
A community-maintained compatibility client and maintenance path for the Web and News API surface used by duck-duck-scrape. It adds typed provider failures, cancellation, deadlines, and proxy control.
- Making MCP context reduction auditable and recoverable.
- Building search and research tools with clear provider boundaries.
- Testing agent tooling with reproducible fixtures and real compatibility checks.
- headroom — Compression and proxy correctness fixes, including preservation of structured content parts.
- CodexPlusPlus — Performance and cross-platform fixes across model catalogs, startup, watchers, and streaming paths.
I write about agent engineering and AI infrastructure on take-a-deep-breath0.com.
Python · TypeScript · Rust · Node.js · MCP