@alexmartinezmGitHub ↗

Alex Martinez

Full-stack engineer who plans with one agent, implements with open-source models via Pi, reviews with frontier models, and keeps a Basic Memory knowledge base as the source of truth.

How this setup has evolved

Version 1 · Latest

Basic Memory as MCP knowledge base

Setup history

Alex's setup

What I build

A B2B SaaS product: .NET backend, React/TypeScript frontend, several repos, and an MCP server that exposes the product to AI clients. AI writes the plans, most of the code and the first review. I decide the design and read every diff before it ships.

How I work

  • Basic Memory is the source of truth. It holds architecture notes, coding standards, product context and ticket decisions, and agents read it over MCP instead of me re-explaining.
  • The main agent turns a ticket and its Figma design into a phased plan, kept outside the repo.
  • Each ticket gets its own git worktree in Orca.
  • For each phase the agent writes a spec: working tree only, no commits, smallest diff, reuse existing components.
  • A Pi agent on an open-source model implements the spec.
  • A Pi agent on a frontier model reviews the uncommitted changes against the plan and the design. A third agent fixes the findings.
  • I read the diff, run builds and tests, check the UI against Figma, and re-verify anything an agent claims. Then I commit in small steps and the agent opens the merge request.

Agents & harnesses

  • Hermes Agent (current): Orchestrator in the terminal on a dedicated coding profile. Plans, writes specs, runs git and builds, dispatches Pi.
  • Pi (current): Runs each implement, review and fix step as a separate process with its own model. Custom roles include glm-implementer, a planner and code reviewers.
  • Orca (current): Manages the worktrees and the terminals where Pi runs.
  • Claude Code, Codex CLI, OpenCode (installed): Claude Code has reviewer and test-writer agents. I used OpenCode for delegation before moving to Pi.

Models

  • Open-source models (current): GLM-5.3-Flash through glm-implementer, plus DeepSeek V4, served from LiteLLM. They implement and fix.
  • Frontier models (current): GPT-6.1 Sol and GPT-6 Astra. They plan, review and sometimes implement or fix (based on custom criteria).

Skills

  • Team conventions (current): Backend, frontend, testing and commit-message rules written as skills so every agent follows them.
  • Plan writing and plan orchestration (current): Turn a ticket into phases and hand them to Pi.
  • MR review (current): Reviews large merge requests in a separate worktree and reads reviewer comments.

MCP connections

  • Basic Memory (current): My knowledge base, searchable by agents. Replaced Obsidian.
  • Figma (current): Design context for UI tickets.
  • Codebase memory (configured): Local code graph for large repos.
  • mysetup (current): Maintains this page.

Moved on from

  • OpenCode for delegation: Pi replaced it.
  • Obsidian as the notes store: Basic Memory replaced it.