Swarf

Invisible, auto-syncing personal storage for any git repo.

You build software and generate byproduct — research docs, design specs, agent skills, scratch notes. Swarf gives this material a durable home alongside any project, without touching the project itself.

The problem

You have files that should live near your code but don't belong in the repo:

• Agent instructions and skills (AGENTS.md, .copilot/skills/)
• Research notes, design docs, open questions
• Security audits, experiment logs, scratch work

Today these files are either untracked and local-only (one rm -rf from gone) or committed to the repo (polluting history, leaking on public repos). Neither is good.

How it works

Swarf creates a swarf/ directory inside your project. Everything in it is automatically excluded from git via .git/info/exclude. A background daemon watches for changes, mirrors them to a central store, and syncs to your configured remote (a git repo or any rclone backend).

my-project/
├── src/
├── swarf/                         ← private storage (invisible to git)
│   ├── docs/                      ← research, design, anything
│   ├── .links/                    ← files projected into the host tree
│   │   └── AGENTS.md
│   └── open-questions.md
├── AGENTS.md → swarf/.links/AGENTS.md   ← symlink (also gitignored)
└── .git/info/exclude              ← managed by swarf

The central store at ~/.local/share/swarf/ mirrors all projects into a single git repo. The daemon commits changes locally, then pushes to your remote. For rclone backends, the entire store is synced — both working files and .git/. This means you can browse your files directly in Google Drive (or whatever backend), and you get full commit history too.

Install

# macOS / Linux (recommended)
brew install swarf

# Via PyPI (persistent install to ~/.local/bin/)
pipx install swarf        # or: uv tool install swarf

# Via Go
go install github.com/mschulkind-oss/swarf@latest

# From source
git clone https://github.com/mschulkind-oss/swarf && cd swarf
just deploy   # builds and copies to ~/.local/bin/

Warning: Don't use pip install swarf inside a virtualenv or uvx swarf. The daemon records the binary's absolute path and breaks when the venv disappears. swarf doctor detects this and warns you.

Quick start

cd ~/projects/my-app
swarf init

On first run, init walks you through setup:

No global config found. Let's set one up.

  Backend [git/rclone] (git): git
  Remote URL (your private backup repo): git@github.com:you/my-swarf.git
✓ Wrote ~/.config/swarf/config.toml
✓ Created central store at ~/.local/share/swarf
✓ Initialized swarf/ for my-app
  Install systemd service for auto-sync? [Y/n] y
✓ Installed systemd service — daemon is running
✓ Daemon is running (PID 12345)

✓ All checks passed.

Next time you run swarf init in another project — no prompts, instant setup.

Rclone backend (Google Drive, Dropbox, S3, etc.)

Set up an rclone remote first:

brew install rclone
rclone config
#   → n (new remote), name: gdrive, type: drive
#   → scope: drive.file (option 2 — only files rclone creates)
#   → auto config: y (opens browser for OAuth)

Then run swarf init, pick rclone, and select your remote from the numbered menu. Swarf defaults to swarf-store as the directory path:

  Backend [git/rclone] (git): rclone

  Pick an rclone remote:

    1. gdrive:

  Enter a number (1-1), or q to quit: 1

  Directory path on gdrive: [swarf-store]:
✓ Remote: gdrive:swarf-store

The store is always a local git repo. Rclone syncs the whole store to the remote — your files are browseable directly in Google Drive (or wherever), and you get full git history via the .git/ directory.

Using swarf

Drop files in

echo "# Design Notes" > swarf/docs/design.md

That's it. The daemon commits and syncs after a 5-second quiet period.

Sweep files into the host tree

Some files need to appear at a specific path in the project (like AGENTS.md in the root). Use sweep to move them into swarf while leaving a symlink:

swarf sweep AGENTS.md
# AGENTS.md → swarf/.links/AGENTS.md

swarf sweep CLAUDE.md .copilot/skills/SKILL.md

Both the original path and swarf/ are automatically gitignored. Symlinks are relative, so they work across machines and inside containers.

To reverse a sweep:

swarf unlink AGENTS.md

Auto-sweep

Configure files to be swept automatically whenever they appear:

# ~/.config/swarf/config.toml
[auto_sweep]
paths = ["AGENTS.md", "CLAUDE.md", ".copilot/skills/"]

The daemon watches for these files and sweeps them on creation.

Check status

swarf status

Shows store state, pending files, remote verification, and daemon health:

Store
  Path           ~/.local/share/swarf
  Backend        git
  Remote         git@github.com:you/my-swarf.git
  Pending        all synced to remote
  Last commit    2 minutes ago — auto: sync 3 files
  Local save     2m ago
  Remote push    2m ago
  Remote sync    verified — local and remote match (a1b2c3d4)

Projects
╭──────────┬─────────────────────┬──────────╮
│ PROJECT  │ PATH                │ STATUS   │
├──────────┼─────────────────────┼──────────┤
│ my-app   │ ~/projects/my-app   │ ✓ ok     │
│ api      │ ~/work/api          │ ✓ ok     │
╰──────────┴─────────────────────┴──────────╯

Daemon: running (PID 12345)

Doctor

swarf doctor checks everything and fixes what it can:

swarf doctor

Doctor handles: missing config, missing store, broken symlinks, absolute symlinks, missing gitignore entries, service installation, and remote reachability. Errors include the exact fix command or config file to edit.

init and doctor share the same engine — the difference is that init creates swarf/ in a new directory, while doctor only checks and repairs.

Commands

Second machine setup

brew install swarf
swarf clone              # clones your store from the configured remote
cd ~/projects/my-app
swarf init               # re-links the project from the store

clone requires global config to exist (with the remote URL). After cloning, run swarf init in each project directory to recreate the local swarf/ directory and symlinks.

Containers and jails

The daemon runs on the host. Containers mount the project directory, so swarf/ comes along. Agents read and write to swarf/ directly — the host daemon picks up changes and syncs.

Inside a container, swarf doctor detects the environment (no global config), skips system checks, and validates project-local state only. sweep and unlink work without the daemon. Symlinks are relative, so they resolve correctly regardless of mount path.

Guides

• Configuration Guide — all config options, file locations, environment variables, and examples
• Roadmap — planned features and distribution channels
• Built-in docs: swarf docs lists all topics (quickstart, architecture, config, sweep, daemon, backends)

License

Apache 2.0