monodev

Coding agents leave files in your repo that don't belong in git: .claude/, AGENTS.md, .cursorrules, scratch scripts, local env files. Commit them and they pollute history; delete them and you lose the context.

A monodev store keeps them outside git. Unapply them for a clean checkout, and apply them again whenever you or an agent need them, hidden from git.

$ git status --short
?? .claude/
?? .cursorrules
?? AGENTS.md
?? debug_helper.py

monodev checkout -n agent-context   # create and activate a store
monodev track --agents              # track the agent paths that exist
monodev track debug_helper.py
monodev save                        # snapshot tracked paths into the store
monodev unapply                     # remove them from the checkout

The files are gone and git status is clean. Bring them back later with:

monodev apply                       # restore them, hidden from git

apply copies the store's files into the working tree and lists them in a managed block in .git/info/exclude, so they stay out of git status while applied. Stores can be pushed to an orphan branch (monodev/persist) so they follow you to other clones without touching main.

Install

macOS and Linux only (why not Windows).

macOS on Apple Silicon (darwin/arm64) — Homebrew:

brew install danieljhkim/tap/monodev

The tap formula points at the darwin/arm64 archive only, so use the release archive instead on any other machine: macOS on Intel (darwin/amd64), Linux amd64, or Linux arm64. Download the archive for your OS and CPU from GitHub Releases, check it against SHA256SUMS, and put monodev on your PATH. Archives also carry shell completions (completions/) and a man page (man/monodev.1); Homebrew installs both for you.

Concepts

• Store: a named set of tracked paths and their contents. By default stores live in <repo>/.monodev/stores/, created on first use with a * .gitignore. Set MONODEV_ROOT=$HOME/.monodev to keep stores outside the repo, which also shares them across worktrees. Stores already in ~/.monodev stay visible either way.
• Workspace: one directory in one checkout, with its active store and applied overlays. Its ID hashes the repo fingerprint and the path from the repo root, so a subdirectory is a separate workspace. See workspace identity.

track --agents covers .claude/, CLAUDE.md, .cursor/, .cursorrules, AGENTS.md, .codex/, .gemini/, .aider* and .github/copilot-instructions.md, skipping any that don't exist.

Agents

Agents can run the loop themselves:

monodev skill init      # teach agents monodev: a SKILL.md, tracked and hidden from git
monodev context         # session start: restore this directory's store, list notes

• skill init writes .claude/skills/monodev/SKILL.md (Claude Code) and/or .agents/skills/monodev/SKILL.md (Codex and other Agent Skills readers), for whichever agent directory exists (--target claude|agents|all to choose). It tracks the skill in the active store. Rerun with --force after upgrading monodev.
• context reports the active store. Without one, it pulls and applies the stores pushed for this directory with push --with-workspace. It never pushes or forces, and exits 0 when it finds nothing.

Notes and scripts are plain files in the store: .agents/notes/lessons.md, one file per session in .agents/notes/sessions/ (so parallel agents never collide), and .agents/scripts/. Keep private or experimental context here. ADRs and gotchas that apply to everyone belong in committed docs.

Commands

monodev status                 # workspace, active store, applied overlays
monodev checkout <store>       # switch active store (-n to create)
monodev track <path>...        # add paths to the active store
monodev save                   # track new files under tracked dirs, then commit
monodev commit --all           # snapshot tracked paths (no discovery)
monodev diff                   # store vs working tree
monodev apply [store...]       # overlay; later stores win conflicts
monodev unapply [--all]        # remove the active store's (or every) overlay
monodev push / pull / sync     # share stores via the persistence branch
monodev eject                  # detach the workspace, keep the files
monodev context [--json]       # restore and list agent context for this directory
monodev skill init / show      # install or print the monodev agent skill
monodev doctor [--fix]         # diagnose and repair local state

push refuses payloads with detected secrets unless you pass --allow-secrets.

Every flag: docs/commands.md.

Guides

• Solo developer in a monorepo
• Sharing stores with a team
• Parallel agents in git worktrees
• Recovering an interrupted apply/unapply/eject

Platform support

Releases ship tar.gz archives for macOS (arm64, amd64) and Linux (amd64, arm64). Each is smoke-tested on a native runner, and the Linux builds also run the real-Git integration suite.

Windows isn't supported. State locking uses flock(2), so the binary doesn't compile there, and a Windows port would need that rewritten first (for example with LockFileEx). No Windows port is planned.

Status

Early development. I built it for my own use, but contributions and design feedback are welcome.

License

MIT