Terminal-first setup for driving hosted and local AI models across Claude Code, Cursor, Codex CLI, and OpenCode. One shared source of truth, symlinked into whatever each tool expects.
One command (gw) drops you into a tmux workspace — Neovim on the left, Claude Code on the right — so editing and AI sessions share one screen. Hand off a task and Claude sorts it into Needs input · Working · Completed while you keep coding.
Background research and tooling comparison that led to this setup: docs/tooling-research.md.
Project conventions (git, snippets, coding style) live in .friday/rules.md, the source of truth symlinked as CLAUDE.md / AGENTS.md / .cursorrules.
gw reads its editor and agent harness from .friday/config.json at the workspace root:
{ "editor": "nvim", "harness": "claude agents" }editor—codeornvim.harness—opencode,claude agents, orcursor-agent.
A second argument (gw <dir> <harness>) or $GW_EDITOR overrides the file; a missing file falls back to nvim / claude agents. Point $GW_CONFIG at a different file to override the location.
.friday/ holds everything shared across harnesses (rules, MCP config, skills, commands) in one place instead of scattered across .claude/, .opencode/, .cursor/, etc. Each harness reads from its own path, so .friday/init symlinks the canonical files into place.
The name is borrowed from Friday, a separate private project: a self-hosted personal AI system — a homelab stack of GenAI tools, content pipelines, and experiments, all running on your own hardware with no cloud dependency by default. It's a nod to Tony Stark's AI assistant F.R.I.D.A.Y. (Female Replacement Intelligent Digital Assistant Youth), the successor to J.A.R.V.I.S. — an always-available, locally-running intelligence that handles the repetitive and creative work so you can focus on what matters.
This repo's .friday/ directory carries the same name because it's the same idea applied to editor tooling: one shared, always-available source of truth instead of config scattered and duplicated across every harness.
.
├── .friday/
│ ├── init # bootstrap script — creates the symlinks below
│ ├── cleanup # removes symlinks .friday/init created
│ ├── rules.md # source of truth for instructions, symlinked as CLAUDE.md / AGENTS.md / .cursorrules
│ ├── config.json # gw editor + agent harness (see "Configuring gw")
│ ├── mcp.json # MCP server definitions, symlinked as .mcp.json
│ ├── skills/ # curated skills surface, symlinked into .claude/skills
│ ├── commands/ # Claude Code-formatted slash commands, symlinked into .claude/commands
│ └── vendor/ # raw upstream git submodules that skills/ symlinks into
└── repos/ # gitignored — clone the repos you work on in here (see below)
bash .friday/initPrompts for which tooling you use (Claude Code, Cursor, Codex CLI/OpenCode) and symlinks accordingly:
| Tool | Link created |
|---|---|
| Claude Code | CLAUDE.md, .mcp.json, .claude/skills, .claude/commands |
| Cursor | .cursorrules |
| Codex CLI | AGENTS.md, .agents/skills/*, ~/.codex/prompts/* |
| OpenCode | AGENTS.md |
Idempotent: re-run any time to fix drifted links. If a real file already exists at a link path, it's backed up to <file>.bak before being replaced. The generated links are per-user setup, not project source, so they're gitignored.
bash .friday/cleanup # remove the symlinks .friday/init createdSkills are Markdown (SKILL.md) directories. Two ways to add one:
-
Own skill: create it directly under
.friday/skills/<name>/SKILL.md. -
Vendored skill (lives inside a third-party monorepo, so a plain submodule would pull in unrelated code):
git submodule add <ssh-url> .friday/vendor/<upstream-repo> ln -s ../vendor/<upstream-repo>/skills/<skill-name> .friday/skills/<skill-name>
vendor/stays an unmodified upstream clone.skills/is the curated, symlinked surface that.friday/initwires into each harness.
Commands are Claude Code formatted Markdown under .friday/commands/, e.g. .friday/commands/friday-snippet.md:
---
description: One-line description shown in the command list
argument-hint: [optional hint for $ARGUMENTS]
---
Command body. Instructions Claude Code follows when you run /friday-snippet.One example already in the repo: /friday-snippet turns a topic into a one-pager under snippets/ (gitignored, never pushed).
.friday/init symlinks .friday/commands → .claude/commands, so any file added here becomes a /name slash command in Claude Code automatically.
Codex CLI gets a best-effort bridge: the same frontmatter (description, argument-hint) and $ARGUMENTS are valid in Codex's custom-prompts format, so .friday/init also links each .friday/commands/*.md into ~/.codex/prompts/, where they're invoked as /prompts:<name>. Custom prompts are user-scoped (not repo-scoped) and deprecated upstream in favor of skills, so treat this as a convenience shim rather than a full first-class Codex equivalent. Cursor and other harnesses don't have a matching format, so commands remain Claude Code-first.
.friday/rules.md is the single source of truth for cross-harness instructions, symlinked as CLAUDE.md, AGENTS.md, and .cursorrules. Edit it once and every harness picks up the change.
repos/ is a gitignored scratch space for cloning the actual repos you work on day to day (e.g. repos/my-app). It's not part of this project's source — it's just a convenient, out-of-the-way place to keep them alongside the shared .friday/ config.
Bootstrap a repo under repos/ against this hub's shared setup instead of creating its own .friday/:
bash .friday/init repos/<name>This wires that repo's CLAUDE.md, .mcp.json, .claude/skills, etc. to point back at this hub's .friday/, so every project you clone into repos/ shares the same rules, skills, and commands.
Harnesses resolve .friday relative to the working directory, so a project can't just point at ~/.friday. Keep a canonical .friday at $HOME instead, and symlink it into each project that wants the shared setup:
# in a project root
ln -s ~/.friday .fridayProjects that need their own overrides keep a real, project-local .friday/ instead of the symlink.