Stack

What I actually run for AI work: models, agent surfaces, and tools, and what each one is for. Not a wishlist.

Last updated November 8, 2026

Current focus

Building multi-agent systems that hold up outside demos: agent architecture, model evaluation, and observability across the coding agents and local inference I actually run day to day.

Models & access

  • Devin (Cognition)

    subscription

    Primary coding subscription for full project builds, refactors, debugging, and architecture work.

  • GPT / Codex (OpenAI)

    subscription

    Codex subscription for cross-model comparison, evaluation tasks, and the codex CLI harness used by local-model-plugin-cc and agent workflow experiments.

  • Local models

    local

    Ollama, LM Studio, and OpenAI-compatible endpoints for local inference. Halogen runs Qwen models; llama-swap style routing and the local-model-plugin-cc review/rescue path support local workflows without cloud dependencies.

Daily drivers

  • Barge

    Cloud runner for coding agents. Dispatches Linear issues and local handoffs to isolated workers that return draft pull requests; runs this site's agent maintenance jobs.

  • Devin CLI

    Primary coding agent. Full project builds, refactoring, debugging, and architecture decisions.

  • Codex CLI and Desktop

    Comparison agent and the broker target for local-model-plugin-cc (Ollama / LM Studio / custom endpoints).

  • OpenCode

    Active daily use. Evaluating how it compares to Devin CLI and Codex for agent-driven workflows.

  • Grok Build CLI

    Hosted Grok review and rescue path via grok-plugin-cc; headless runs with explicit safety gates.

  • Mission Control

    Unified usage and runtime dashboard across Codex, Hermes, and related lab sources (sessions, tokens, failures, and machine load).