I build tools that help AI agents understand code, debug failures, and work across real systems.
Senior Systems Engineer, GenAI Operations. Based in Madrid. My recent work spans execution recording, time-travel debuggers, agent harnesses, MCP integrations, and workflow automation. I work across TypeScript, Python, Go, and Java.
Pyokka · Python and Tracelight · JavaScript / TypeScript
Two projects built around the same question: what did this program actually do? They expose live values, recorded execution, and debugging through VS Code, a browser interface, a CLI, and MCP. An agent can inspect a run, follow a value's history, or pause the program and investigate alongside you.
Pyokka in action: an agent drives the debugger through function breakpoints and stepping while the editor shows the paused values. Watch the video.
Pyokka guide · Tracelight runtime internals · Tracelight wire protocol
Rewalk records what someone does and says while using a web page, then connects spoken complaints to DOM changes and React components. Coding agents get the relevant browser evidence, bug reports, and replays. Watch the demo.
agent-cli-scaffold / agtcli is a Python runtime and scaffold for CLIs agents can drive. Generated tools start with structured output, typed exit codes, credential handling, request guards, and contract tests. A checker makes the interface rules inspectable.
har2cli turns an authenticated browser network capture into a redacted HTTP contract, tests a selected GET request, and hands it to agtcli to generate a CLI. It keeps credentials separate from exported contracts and bounds the requests used to probe authentication.
| Project | What it does |
|---|---|
| WorkDone | Coordinates Claude Code, Codex, and Cursor agents across machines through ChatGPT and MCP. Voice interaction, persistent sessions, agent events, and approval handling. |
| Farrier | Generates repository-specific agent harnesses with rules, tested guard hooks, and verification commands. Includes session analysis and evaluations that distinguish observed results from unproven benefits. |
| Claude Agent SDK for n8n | Runs agent tasks inside self-hosted workflows, with multiple providers, memory, human approvals, MCP tools, structured output, and persistence. |
| Mercadona CLI | Gives agents a structured interface to a real grocery service: product search, prices, and cart management. Written in Go and distributed as a standalone binary and through npm. |
More tools for agent context, integrations, and verification
Debugging and agent context. PyCharm Debugger MCP exposes the IDE's debugger to agents. sxr searches Claude Code and Codex session histories; transcript-reader provides a local reader for Claude Code sessions. sk loads skills on demand instead of putting every skill into the initial context.
orly gives agents access to O'Reilly Answers with book citations.
Checks and integrations. konpy enforces structural conventions in Python repositories. CodeMie for n8n adds SSO proxy authentication to the Claude Agent SDK node. Raycast Prompter puts editable LLM writing prompts behind a keyboard shortcut.
I also build Herdr plugins for agent notifications, agent lifecycle management, local ports, and tunnels.