GitFerry MCP server

gitferry

v0.4.2 is released. npm package · v0.4.2 release and downloads · MCP Registry

Portable git-change bundles for air-gapped repos. Pack your unpushed commits, staged/unstaged changes, and untracked files into a single shareable file on one machine — carry it across the air gap (USB, secure file transfer, email, or QR codes across a webcam) — validate and apply it to the same repo on another machine. No shared remote, no network, no sync required.

machine A (has your changes)          machine B (same repo, no sync)
┌─────────────────────────┐           ┌─────────────────────────┐
│ $ gitferry pack         │  ──────▶  │ $ gitferry apply        │
│   → repo-main-a1b2c3d   │  USB /    │   ✓ checksums verified  │
│     .gitferry           │  scp /    │   ✓ same repo confirmed │
│                         │  email    │   ✓ 2 commits replayed  │
│  commits + worktree     │           │   ✓ staged/unstaged     │
│  + untracked, checksummed│          │     state restored      │
└─────────────────────────┘           └─────────────────────────┘

See SCENARIOS.md for every situation this is built for — air-gapped environments, client-site work, WIP handoffs, offline review, and more. See INSTALL.md for offline installation on other devices.

Quick MCP install (Claude Code): claude mcp add --transport stdio gitferry -- npx --yes --package=gitferry@0.4.2 gitferry-mcp, then restart your client and confirm the pack_changes, inspect_bundle, and apply_bundle tools appear. Codex, VS Code, JSON config, and offline setup are in MCP server below.

Why not just git format-patch / git bundle?

You can — gitferry uses them under the hood. What raw git doesn't give you:

• Untracked files travel too (respecting .gitignore) — git diff never sees them
• Staged vs unstaged state is preserved — your index looks the same on the other side
• Validation before anything touches the repo: sha256 checksums of every file, same-repository fingerprint check (root commits), base-commit check, clean-tree check
• One command each way instead of a multi-step format-patch/diff/am/apply choreography
• Binary-safe (--binary patches for tracked files, verbatim bytes for untracked)
• Plain-git fallback: the bundle is a zip of standard git patches — pack with --portable and a machine with only git + unzip applies it by hand (instructions are inside every bundle; the compact default needs node's one-liner to unwrap first)

Install

Requires Node.js 22 or newer and Git 2.32 or newer. Install the published CLI and MCP server from npm:

npm install -g gitferry@0.4.2
gitferry --version

For MCP-only use, skip the global install and use the pinned npx configuration in MCP server.

Installation and MCP registration are separate steps. Installing the npm package makes gitferry and gitferry-mcp available in your terminal. Registering the server tells Codex, Claude Code, or another MCP client how to launch it. With npx, the client downloads the pinned package on first launch, so a global install is optional. Reading this README or installing the package alone does not register the server.

Closed environments (no npm)

Download gitferry.cjs, gitferry-mcp.cjs, and SHA256SUMS from the public v0.4.2 release. These standalone files include their JavaScript dependencies; transfer them to the offline machine with Node and Git. See INSTALL.md for checksum verification and setup.

To build the same standalone executables from a source checkout:

npm run bundle                 # → dist/gitferry.cjs + dist/gitferry-mcp.cjs
node dist/gitferry.cjs pack    # runs without any install or node_modules

Usage

On the machine with changes

gitferry pack                          # asks what should travel (Enter = uncommitted); base = merge-base with upstream
gitferry pack --only staged            # skip the question: all | uncommitted | staged | unstaged | committed | repo
gitferry pack --only repo              # whole-repo snapshot: every tracked file as-is — seed/sync another machine
gitferry pack --base origin/main       # explicit base
gitferry pack -o changes.gitferry --no-untracked

The interactive picker only appears in a terminal; in scripts (no TTY) pack takes everything unless --only or --no-* flags say otherwise. Selections mean: all = commits + staged + unstaged + untracked · uncommitted = everything except commits · staged = index edits only · unstaged = worktree edits + untracked · committed = unpushed commits only · repo = a snapshot of every tracked file as it is on disk (no history) — apply seeds a bare directory or syncs an existing checkout, skipping identical files and touching drifted ones only with --force.

On the receiving machine

gitferry inspect changes.gitferry      # verify + preview, touches nothing
gitferry apply changes.gitferry --dry-run
gitferry apply changes.gitferry

No transfer channel at all? Beam it over QR

When even USB is off the table, the screen itself is the channel: one machine plays the bundle as a looping sequence of QR codes, the other watches with any webcam.

# machine A (sender)
gitferry show changes.gitferry             # QR loop right in the terminal
gitferry show changes.gitferry --browser   # bigger + denser, in a local browser page
gitferry show changes.gitferry --color     # experimental: 3x density, 3 chunks per image in R/G/B

# machine B (receiver) — point its webcam at A's screen
gitferry read                              # opens a local camera page, reassembles the bundle
gitferry apply changes.gitferry

Every QR frame carries its chunk number / total plus a transfer id, so the receiver knows exactly which pieces it still needs — it shows the missing numbers live and keeps collecting as the loop repeats (order doesn't matter; missed frames come around again). The file is only written once every numbered chunk arrived and the whole-file sha256 matches, so a completed read is proof nothing was skipped or corrupted. Both pages live on 127.0.0.1; no network between the machines is used or needed.

Point the sender's webcam back at the receiver and the loop gets smart. The read page displays a small progress QR of its own (tiny payload, high error correction, so it scans easily). If the show --browser machine also has a camera watching the receiver's screen, then after each loop the sender re-shows only the chunks that are still missing, dropping from the first-loop rate (default 4 fps) to a slower retry rate (default 1 fps, --retry-fps) so the stragglers are easy to catch — and stops automatically once the receiver's code says complete. No camera on the sender's side? Everything simply loops at full speed like before.

Throughput is roughly chunk × fps: defaults give ~0.3 KB/s in the terminal, ~0.6 KB/s in the browser, ~1.8 KB/s with --color — meant for patch-sized bundles, not megabyte piles (pack --no-untracked helps keep them small). The default is deliberately a coarse, easy-scan code (150 bytes ≈ QR version 8, 49×49 modules): webcams reading another screen fail on fine modules long before they fail on frame count. Raise --chunk (up to 2900) if your camera and screens are good.

Applying

Apply order: commits (git am -3, full author/date/message preserved) → staged patch (index + worktree) → unstaged patch (worktree only) → untracked files (never overwrites existing files unless --force).

On conflicts you get standard git conflict markers and clear instructions (git am --continue / --abort, .rej files as last resort) — application is not transactional, so review the error and recovery instructions before continuing.

Flags

Flag	Command	Effect
--base <ref>	pack	commit the bundle applies onto
--only <what>	pack	all / uncommitted / staged / unstaged / committed / repo — non-interactive selection
--no-commits / --no-worktree / --no-untracked	pack	trim what's captured (combinable with --only)
--portable	pack	plain-zip bundle (~30% bigger) for targets with only git + unzip
--dry-run	apply	validate + show the plan, change nothing
--force	apply	override same-repo, clean-tree, and overwrite guards
-C <dir>	pack/apply	run against another directory
--chunk <bytes> / --fps <n>	show	QR payload size (default 150 ≈ version-8 code; raise for speed, lower for even easier scanning) and first-loop display speed
--retry-fps <n>	show	slower rate for re-loops of missing chunks once the receiver's progress QR is seen (default 1)
--browser / --color	show	browser-page display; experimental 3× density via R/G/B channels
--no-open / --port <n>	show/read	don't auto-open the local page / pin its port
-o <file> / --force	read	output path / overwrite an existing file

MCP server

GitFerry runs locally over stdio. Your MCP client starts the process, and it works with repositories on that same machine. No GitFerry account, API key, hosted service, or source upload is required. URL-only MCP clients need a separate bridge and cannot connect directly to this local server.

Register with the CLI for your client (requires Codex CLI or Claude Code, respectively):

# Codex
codex mcp add gitferry -- npx --yes --package=gitferry@0.4.2 gitferry-mcp
codex mcp get gitferry

# Claude Code
claude mcp add --transport stdio gitferry -- npx --yes --package=gitferry@0.4.2 gitferry-mcp
claude mcp get gitferry

Choose your client's registration command, reconnect or start a new session, and verify that pack_changes, inspect_bundle, and apply_bundle are available. mcp get confirms the saved configuration; tool discovery in the connected client confirms the server is working. See Codex setup for the TOML alternative and agent-assisted installation to delegate setup.

Use this configuration for the published 0.4.2 package in clients that accept mcpServers JSON. The explicit executable matters because this package supplies both a CLI and an MCP server:

{
  "mcpServers": {
    "gitferry": {
      "command": "npx",
      "args": ["--yes", "--package=gitferry@0.4.2", "gitferry-mcp"]
    }
  }
}

For a source checkout, build with npm ci && npm run bundle and use:

{
  "mcpServers": {
    "gitferry": {
      "command": "node",
      "args": ["/absolute/path/to/gitferry/dist/gitferry-mcp.cjs"]
    }
  }
}

If you installed the package globally, the server command is gitferry-mcp. For a downloaded offline bundle, point node directly at /absolute/path/to/gitferry-mcp.cjs. See client setup for Codex, Claude Code, and VS Code examples and Windows launch instructions.

Tool	Action	Key arguments
pack_changes	Writes a portable bundle from local changes	cwd, optional output, base, only, portable and include flags
inspect_bundle	Reads and validates a bundle without changing a repository	bundlePath
apply_bundle	Previews or applies a bundle to a local target	cwd, bundlePath, dryRun, force, includeUntracked

Use absolute paths. Start with: “Pack the uncommitted changes in /absolute/path/to/repo into /absolute/path/to/transfer.gitferry.” On the receiving machine: “Inspect this bundle, then dry-run it against /absolute/path/to/repo and show the plan before applying.”

only accepts all, uncommitted, staged, unstaged, committed, or repo. Snapshot mode (repo) carries current tracked-file contents rather than history; it can seed an empty directory. The default without a selection includes all change categories, matching non-interactive CLI use. Tool annotations describe side effects to the client; they do not enforce approvals or restrict filesystem access.

See INSTALL.md for setup and offline transfer, DEMO.md for a two-repository walkthrough, and PUBLISHING.md for maintainer release steps. Published registry identity: io.github.mayank1307/gitferry.

Agent-assisted installation

Paste this into Codex or Claude Code with terminal and configuration access:

Set up GitFerry v0.4.2 as a local stdio MCP server in this client using https://github.com/mayank1307/gitferry/blob/main/INSTALL.md. Check Node.js 22+ and Git 2.32+. Check for an existing GitFerry registration, preserve other servers, and register the pinned npx command for this client. A global npm install is optional. Verify the saved configuration, then verify that pack_changes, inspect_bundle, and apply_bundle are available. If a restart or new session is needed, tell me what to do and verify tool discovery afterward.

The agent can perform setup when asked and given the required access. Package download happens automatically through npx after registration; client approvals and reconnection follow the client's normal behavior.

Bundle format

A .gitferry file is a zip with this layout — by default brotli-wrapped (GFBR magic + brotli stream of a stored zip, ~30% smaller than a deflate zip; pack --portable writes a plain zip instead, and inspect/apply auto-detect both):

manifest.json            # versioned manifest: repo fingerprint, base/head shas,
                         # content listing, sha256 of every entry
commits/00NN-*.patch     # git format-patch output (mbox — author, date, message)
worktree/staged.patch    # git diff --binary --cached
worktree/unstaged.patch  # git diff --binary
untracked/<paths>        # verbatim file bytes
README.txt               # human instructions for manual apply with plain git

Everything inside is standard-git-consumable by design; gitferry adds integrity checking and one-command UX, never a proprietary encoding. For a target with only git + unzip, pack with --portable; the compact default can still be unwrapped by any machine with node (the one-liner is in the bundle's own README.txt).

Notes & limitations

• Merge commits in the packed range are flattened by format-patch — pack warns; prefer linear branches or pick a different --base
• Replayed commits get new shas (same content/author, new committer) — normal for git am
• Change bundles require shared repository history (validated via root-commit fingerprint); snapshot bundles can also seed an empty directory.
• Requires git ≥ 2.32 on the applying machine

Safety and trust

Bundles contain source code, commit messages, and potentially untracked secrets. Review the selection and destination before sharing. .gitignore excludes ignored untracked files; it does not remove secrets already tracked in Git. Checksums detect corruption, not the identity or trustworthiness of the sender. Only apply bundles from sources you trust.

The server has the filesystem privileges of the process that starts it. Configure it only in trusted clients and run it as your normal user. force bypasses repository, dirty-tree, and overwrite guards. See SECURITY.md for enforced limits, symlink behavior, and concurrency boundaries.

Development

See repository architecture for the source layout and MCP deployment for local, npm, and Docker setup. Copyable client configurations live in examples/mcp.

npm ci
npm run build     # tsc → dist/
npm test          # vitest: round-trip e2e on real temp repos + guard tests
npm run bundle    # fresh single-file CLI + MCP server
npm run release:check  # versions, build, tests, and isolated package smoke test