I build deterministic, verifiable tooling for AI coding agents: small tools that return measured results instead of model guesses, and that say "not validated" when they cannot know. Physics student, so I like claims that come with evidence.

Español: construyo herramientas deterministas y verificables para agentes de IA; los repositorios están documentados en inglés y español.

_{Rendered with opus-studio: every frame is a pure function of time. Each repository below has its own demo.}

Agent safety and credentials

• agent-credential-broker - lets an agent perform authenticated actions without ever holding the API key or password (local encrypted vault, scoped tasks, quotas, metadata audit). Experimental, not independently audited.
• enlace - connects agents to applications by researching the interface first, then planning and executing exact operations (SDK, CLI, MCP). Experimental.

Observability and sessions

• agent-observer - provider-neutral local logging of agent metadata and evidence-based operational reviews.
• pi-session-workspace - session navigation, checkpoints and a local mailbox inside the Pi terminal agent.
• claude-agent-continuity - keeps the work of a cut-off Claude Code subagent: safe WIP commits per git worktree, status, resume prompts and a hook.

Debugging and deterministic guardrails

• rubber-duck - makes a coding agent debug like a programmer explaining code to a rubber duck: it narrates the execution path as claims, checks each with a command, and may only edit after a claim is refuted. Claude Code plugin.
• rosetta-guards - detects identifiers lost when an LLM compacts context, plus a controlled-vocabulary linter. No LLM, no dependencies.
• realign - checks declared relationships between project files and repairs only derived text blocks, with verifiable evidence.

3D and CAD for agents

• agentcad - deterministic checks for OpenSCAD/STL designs (interference, printability, spec-as-code). A digital result is never labelled physical.

Programmatic motion graphics

• opus-studio - code-driven film production engine for autonomous agents: pure seek(t) rendering, direction lint, fragment-based editing.

• Deterministic. Checks are plain code, so the same input gives the same answer and a gate can rely on it.
• Provider-neutral. No dependence on one model, agent framework, terminal or account.
• Zero or minimal dependencies. Standard library wherever possible.
• Evidence over claims. Limits are documented, unknowns are reported as unknown, and experimental status is stated up front.

JavaScript / Node.js, Python, OpenSCAD, Linux, git, MCP.

Open an issue or discussion on any of the repositories above.