How I Actually Work With AI (OpenCode, Antigravity, Claude, and Codex)

I have wanted to document my AI workflow for a while.

I switch between these tools every day, but I did not have clear numbers on which models and workflows I rely on most.

So I checked config files, logs, and session databases across OpenCode, Antigravity, Claude, and Codex.

AI tools I use

Before going deep, here is the short version:

• OpenCode CLI: Where I run larger coding tasks, agent workflows, and verification.
• Antigravity editor: My main in-editor environment while I write and review code.
• Claude Code (inside Antigravity): My most-used tool for day-to-day work tasks, especially docs and implementation support.
• Codex CLI: What I use when I need focused, high-reasoning coding sessions.

I also rely on shared skills and MCP servers across these tools.

What I checked

I pulled data from these local paths:

• OpenCode: ~/.config/opencode/, ~/.local/share/opencode/, ~/.local/share/opencode/opencode.db
• Antigravity: ~/Library/Application Support/Antigravity/, ~/.gemini/antigravity/
• Claude: ~/.claude.json, ~/.claude/settings.json, ~/.config/claude-mcp/mcp.json
• Codex: ~/.codex/config.toml, ~/.codex/sessions/
• Skills: ~/.agents/skills/, ~/.claude/skills/

Claude Code is where most work happens

After checking the local data, this became very clear.

From ~/.claude/stats-cache.json:

• totalSessions: 30
• totalMessages: 9769
• Date range covered: 2026-01-06 to 2026-03-15

For work-related tasks, Claude Code is not just part of my workflow. It is the main one.

OpenCode setup and usage

My primary OpenCode config is in ~/.config/opencode/opencode.json.

Current defaults:

• model: google/claude-opus-4-5-thinking
• small_model: google/gemini-2.5-flash
• Plugins:
  • oh-my-opencode@latest
  • opencode-openai-codex-auth
  • opencode-antigravity-auth@1.6.0

OpenCode plugins I run

These are the OpenCode plugins configured in ~/.config/opencode/opencode.json:

• oh-my-opencode@latest
• opencode-openai-codex-auth
• opencode-antigravity-auth@1.6.0

I also use Oh My OpenCode routing in ~/.config/opencode/oh-my-opencode.json.

Notable mappings there:

• sisyphus, oracle, prometheus, metis: GPT 5.2 variants
• explore: opencode/gpt-5-nano
• librarian: opencode/big-pickle
• ultrabrain category: openai/gpt-5.3-codex

GSD (Get Shit Done)

GSD is the workflow I use when I want AI work to be structured and repeatable.

On my machine, it lives in ~/.config/opencode/get-shit-done/ (version 1.25.1), and the command definitions are in ~/.config/opencode/command/gsd-*.md.

GSD’s core loop is:

• /gsd-new-project
• /gsd-plan-phase
• /gsd-execute-phase

The commands I use the most:

• /gsd-quick for small tasks when I already know what to do
• /gsd-progress to get status and route to the next action
• /gsd-verify-work to do conversational UAT (one test at a time)

Most-used models in OpenCode (from local DB)

From ~/.local/share/opencode/opencode.db, grouped by assistant provider/model:

Provider	Model	Count
anthropic	claude-opus-4-5	1045
openai	gpt-5.1-codex-max-high	500
google	claude-opus-4-5-thinking	287
openai	gpt-5.1-codex-max-xhigh	194
google	antigravity-gemini-3-pro	178
openai	gpt-5.3-codex	156
google	gemini-3-pro-high	130

Antigravity-routed models inside OpenCode history:

• antigravity-gemini-3-pro: 178
• antigravity-claude-opus-4-5-thinking: 79

The practical takeaway is simple: I do not run one model all day. I route by task.

Antigravity + Claude Code

Antigravity editor settings are in ~/Library/Application Support/Antigravity/User/settings.json.

Current Claude Code editor preferences there:

• claudeCode.preferredLocation: panel
• claudeCode.selectedModel: default

From Antigravity logs, I found these useful model signals:

1. Launch events in Claude VSCode.log commonly start with model:"default" or model:"sonnet".
2. Explicit set-model events include:
   • haiku
   • sonnet
   • opus
   • claude-opus-4-6

Claude setup, skills, and MCPs

From ~/.claude/settings.json:

• Global model is set to opus.

From my Claude environment:

• Enabled plugins include:
  • superpowers
  • context7
  • greptile
  • gopls-lsp
  • typescript-lsp
  • code-review
  • modern-go-guidelines

Custom skills are installed in both ~/.claude/skills/ and ~/.agents/skills/:

• find-skills
• postgres
• mysql
• technical-writer

From ~/.claude.json, technical-writer is currently my most-used custom skill.

Claude MCP server definitions appear in:

• ~/.claude.json
• ~/.config/claude-mcp/mcp.json

MCP servers defined there:

• magic
• excalidraw
• context7
• exa
• css (via css-mcp)

Codex setup

From ~/.codex/config.toml:

• model = "gpt-5.3-codex"
• model_reasoning_effort = "high"
• personality = "pragmatic"

From local session artifacts in ~/.codex/sessions/, historical model density is strongest around:

• gpt-5.1-codex-max
• gpt-5.2-codex
• then gpt-5.3-codex

So my current Codex default is gpt-5.3-codex, while history still carries more 5.1/5.2 runs.

MCP map across tools

Tool	MCP signal
OpenCode	Runtime logs show websearch, context7, grep_app, magic
Claude	Config includes magic, excalidraw, context7, exa, css
Antigravity	-
Codex	-

How I work day to day

This is my normal loop:

1. Start in Antigravity for direct editing.
2. Use Claude Code for in-editor iteration and model switching.
3. Move to OpenCode when I need orchestration, parallel search, and stronger execution structure.
4. Use Codex CLI for focused high-reasoning coding sessions.
5. Use skills and MCPs as shared capability across all of the above.

What changed after this audit

This review was useful for three reasons:

• I now have measured usage, not assumptions.
• I can clearly see where defaults and real behavior differ.
• I know what to tune next: model defaults, skill loading, and MCP consistency.

If you use multiple AI tools daily, I recommend doing this once.

Your machine history is usually more accurate than your memory.