claude-vigil-mcp
An MCP server for creating named checkpoints and snapshots of projects, with selective file restore, content diffs, and artifact preservation, enabling safe recovery from external file changes in Claude Code.
README
<img align="right" src="claude-vigil.svg" alt="claude-vigil-mcp" width="220">
claude-vigil-mcp
An Model Context Protocol (MCP) server for checkpoint, snapshot, and file recovery in Claude Code. Perfect snapshots, selective restore, bash safety net, and honest disk management.
<br clear="right">

Every AI coding tool tracks file edits made through its own editor, but none of them track file changes made externally: bash commands (rm, mv, sed -i), Python scripts, build tools, or any process that modifies files outside the editor's API. When those changes go wrong, there's nothing to rewind to. Claude Code's built-in /rewind has additional gaps -- external changes are invisible (#6413, #10077), rewind is all-or-nothing with no selective file restore, timestamps only with no named checkpoints, and reliability bugs (#21608, #18516).
install
Requirements:
From shell:
claude mcp add claude-vigil-mcp -- npx claude-vigil-mcp
From inside Claude (restart required):
Add this to our global mcp config: npx claude-vigil-mcp
Install this mcp: https://github.com/Vvkmnn/claude-vigil-mcp
From any manually configurable mcp.json: (Cursor, Windsurf, etc.)
{
"mcpServers": {
"claude-vigil-mcp": {
"command": "npx",
"args": ["claude-vigil-mcp"],
"env": {}
}
}
}
There is no npm install required -- no external databases, no indexing, only Node.js built-ins for crypto, compression, and filesystem.
However, if npx resolves the wrong package, you can force resolution with:
npm install -g claude-vigil-mcp
skill
Optionally, install the skill to teach Claude when to proactively checkpoint before risky work:
npx skills add Vvkmnn/claude-vigil-mcp --skill claude-vigil --global
# Optional: add --yes to skip interactive prompt and install to all agents
This makes Claude automatically save checkpoints before destructive bash commands, risky refactors, or context compaction. The MCP works without the skill, but the skill improves discoverability.
plugin
For automatic checkpointing with hooks and commands, install from the claude-emporium marketplace:
/plugin marketplace add Vvkmnn/claude-emporium
/plugin install claude-vigil@claude-emporium
The claude-vigil plugin provides:
Hooks (background, zero-latency):
PreToolUse (Bash)- auto-quicksave before destructive commands (rm,mv,sed -i,git checkout,git reset)PreCompact- auto-checkpoint before context compaction, both manual (/compact) and automaticStop- auto-checkpoint after Claude finishes a response that included file editsPostToolUse (Write|Edit)- checkpoint after file modificationsSessionEnd- last-chance checkpoint when the session terminates
Command: /checkpoint <save|list|diff|restore|delete>
Requires the MCP server installed first. See the emporium for other Claude Code plugins and MCPs.
features
5 tools. Perfect snapshots. Content diffs. Safe restores with artifact preservation.
vigil_save
Create a named checkpoint of the entire project. Optional description for annotation. If slots are full, Claude asks the user whether to delete an existing checkpoint or increase capacity.
🏺 ┏━ saved "before-refactor" ━━ 47 files · 4.2 MB ━━ vigil: 2/3 | quicksave: 8m ago | 4.2 MB
┗ skipped: node_modules, dist, .next
First save auto-detects derived directories from .gitignore and creates .vigilignore:
🏺 ┏━ saved "v1.0" ━━ 47 files · 4.1 MB ━━ vigil: 1/3 | quicksave: none | 4.1 MB
┃ skipped: node_modules, dist, .next
┃ first save -- confirm these exclusions look correct
┗ edit .claude/vigil/.vigilignore to adjust
When slots are full:
🏺 ┏━ 3/3 full -- ask the user before proceeding ━━ vigil: 3/3 | quicksave: 2m ago | 8.7 MB
┃ v1.0 (2h ago) · before-refactor (45m ago) · experiment (5m ago)
┗ ASK the user: delete one with vigil_delete, or increase capacity with max_checkpoints?
vigil_list
Browse checkpoints with descriptions. With name: drill into that checkpoint's files. With glob: filter files by pattern.
🏺 ┏━ 2 checkpoints ━━ vigil: 2/3 | quicksave: 3m ago | 8.7 MB
┃ v1.0 2h ago 47 files
┃ Initial stable release
┃ before-refactor 45m ago 47 files
┃ Snapshot before risky auth changes
┗ ~quicksave 3m ago
Drill into a checkpoint with glob filtering:
vigil_list name="v1.0" glob="src/auth/**"
🏺 ┏━ v1.0 ━━ 3 of 47 files matching src/auth/** ━━ vigil: 2/3 | quicksave: 3m ago | 8.7 MB
┃ src/auth/index.ts
┃ src/auth/middleware.ts
┗ src/auth/types.ts
vigil_diff
Search and investigate previous versions of your codebase. Compare a checkpoint against the current working directory with full unified diffs, compare two checkpoints against each other, retrieve any file's content from any checkpoint, or search for a string across all checkpoints.
Summary of changes:
vigil_diff name="before-refactor" summary=true
🏺 ┏━ 3 changes vs before-refactor ━━ vigil: 2/3 | quicksave: 3m ago | 8.7 MB
┃ modified src/auth.ts (+8 -2)
┃ modified src/middleware/validate.ts (+3 -1)
┗ added src/services/oauth.ts
Full unified diffs:
vigil_diff name="before-refactor"
🏺 ┏━ 3 changes vs before-refactor ━━ vigil: 2/3 | quicksave: 3m ago | 8.7 MB
┃ modified src/auth.ts (+8 -2)
┃ modified src/middleware/validate.ts (+3 -1)
┗ added src/services/oauth.ts
━━ src/auth.ts ━━
--- a/src/auth.ts
+++ b/src/auth.ts
@@ -12,6 +12,8 @@
import { validateToken } from './utils';
-function authenticate(req: Request) {
+function authenticate(req: Request, options?: AuthOptions) {
+ if (options?.skipValidation) return true;
const token = req.headers.authorization;
Retrieve a single file from a checkpoint:
vigil_diff name="v1.0" file="src/auth.ts"
🏺 ━━ src/auth.ts from v1.0 ━━
import { validateToken } from './utils';
function authenticate(req: Request) {
const token = req.headers.authorization;
...
━━ diff vs current ━━
--- a/src/auth.ts
+++ b/src/auth.ts
@@ -12,6 +12,8 @@
-function authenticate(req: Request) {
+function authenticate(req: Request, options?: AuthOptions) {
Compare two checkpoints:
vigil_diff name="v1.0" against="before-refactor"
Shows unified diffs between the two checkpoint states -- no working directory involved.
Search across all checkpoints:
vigil_diff name="*" file="src/auth.ts" search="validateToken"
🏺 ┏━ "validateToken" in src/auth.ts ━━ 2 checkpoints ━━ vigil: 2/3 | quicksave: 3m ago | 8.7 MB
┃ v1.0 (2h ago)
┃ import { validateToken } from './utils';
┃ before-refactor (45m ago)
┗ import { validateToken } from './utils';
vigil_restore
Restore the project to a checkpoint state. Quicksaves current state first (undo with vigil_restore name="~quicksave"). Displaced files -- both modified and newly created since the checkpoint -- are preserved in .claude/vigil/artifacts/ so nothing is ever lost. For individual file restores, use vigil_diff to retrieve file content, then apply with Edit.
vigil_restore name="v1.0"
🏺 ┏━ restored from "v1.0" ━━ 47 files ━━ vigil: 2/3 | quicksave: just now | 8.7 MB
┃ preserved 3 displaced files in .claude/vigil/artifacts/restored_v1.0_20260219_143022/
┃ modified: src/auth.ts (current version saved)
┃ modified: src/middleware/validate.ts (current version saved)
┃ new: src/services/oauth.ts (moved, not in checkpoint)
┃ review .claude/vigil/artifacts/restored_v1.0_20260219_143022/ -- delete when no longer needed
┃ previous state also quicksaved (use ~quicksave to undo)
┃ not restored (derived): node_modules, dist
┗ rebuild these before running the project
vigil_delete
Delete a checkpoint and reclaim disk space. GC removes unreferenced objects. Use all=true to delete everything.
vigil_delete name="v1.0"
🏺 ━━ deleted v1.0 ━━ reclaimed 241 MB (3,412 objects) ━━ vigil: 1/3 | quicksave: 3m ago | 4.5 MB
methodology
How claude-vigil-mcp stores checkpoints:
🏺 claude-vigil-mcp
━━━━━━━━━━━━━━━━━━━
Claude calls tool
vigil_save
│
▼
┌─────────────────┐
│ spawn worker │ <5ms, returns immediately
│ (detached) │
└────────┬────────┘
│
┌──────────┴──────────┐
│ background worker │
│ │
│ ┌───────────────┐ │
│ │ walk project │ │ source files only (skips derived dirs)
│ └───────┬───────┘ │
│ │ │
│ ┌───────▼───────┐ │
│ │ hash (SHA-256)│ │ same content = same hash
│ └───────┬───────┘ │
│ │ │
│ ┌───────▼───────┐ │
│ │ gzip + store │ │ dedup: skip if exists
│ └───────┬───────┘ │
│ │ │
│ ┌───────▼───────┐ │
│ │update manifest│ │ {path → hash} per checkpoint
│ └───────────────┘ │
└─────────────────────┘
┌─────────────────────────────────────────┐
│ .claude/vigil/ │
│ ├── manifest.json checkpoints + meta│
│ ├── objects/ │
│ │ ├── ab/cdef01...gz gzipped file │
│ │ ├── f3/981a02...gz gzipped file │
│ │ └── ... (deduped) │
│ └── artifacts/ │
│ └── restored_v1.0_20260219_.../ │
│ ├── src/auth.ts (modified) │
│ └── src/new.ts (new file) │
└─────────────────────────────────────────┘
3 named slots + 1 rotating quicksave
Every response: "vigil: 2/3 | quicksave: 3m ago | 287 MB"
RESTORE (only sync operation):
vigil_restore("v1.0")
│
├── quicksave current state (overwrite previous)
│
├── preserve displaced files in artifacts/
│ ├── modified files → copied to artifacts
│ └── new files → moved to artifacts
│
├── read manifest → get {path → hash}
│
├── for each file: gunzip object → write to project
│
└── done: bit-identical working directory
Storage: Content-addressable storage (SHA-256 + gzip). Same file across checkpoints = stored once. Binary files included -- a restored checkpoint is bit-identical to the original.
Performance: Background worker via spawn(detached). MCP tool returns in <5ms. Worker runs independently. Only vigil_restore is synchronous (must write files before Claude proceeds).
Disk honesty: Every tool response shows vigil: 2/3 | quicksave: 3m ago | 273 MB. No hidden costs. 3 checkpoint slots by default. .vigilignore for excluding paths you don't need.
Artifact preservation: On restore, files that would be overwritten or lost (modified since checkpoint, or newly created) are preserved in .claude/vigil/artifacts/. Nothing is ever deleted -- you can always recover displaced work.
v1.0 v1.1 v1.2 objects/
━━━━ ━━━━ ━━━━ ━━━━━━━━━━━━━━━━
src/index.ts ab3f ══════ ab3f ══════ ab3f → ab/3f01a2...gz
src/auth.ts f981 e904 ══════ e904 → f9/81b3c4...gz
e9/04f7a8...gz
src/server.ts 2bc4 ══════ 2bc4 ══════ 2bc4 → 2b/c4d5e6...gz
src/utils.ts 7de1 ══════ 7de1 ══════ 7de1 → 7d/e1f2a0...gz
src/config.ts 4aa2 ══════ 4aa2 d71c → 4a/a2b1c3...gz
d7/1c45e8...gz
──── ──── ────
new objects: 5 1 1 = 7 (not 15)
══════ same SHA-256 across checkpoints -- stored once, referenced many
save v1.0 ██████████████████████████████████████████████████ 100 new
save v1.1 ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░████ 8 new
save v1.2 ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░██ 3 new
──────────────────────────────────────────────────
111 objects store 300 file-versions 2.7× dedup
delete v1.0 gc sweep → 62 orphans freed, 49 shared retained
░░░░░░░░░░░░░░░░░░░░░░░░░ 4.1× dedup
████ new object stored ░░░░ deduped (hash exists, write skipped)
Core techniques:
- content-addressable storage (
storeObject): SHA-256 hash →existsSynccheck → skip or gzip + write. Same content = same address = automatic dedup. - mark-and-sweep GC (
gcObjects): Union referenced hashes across all checkpoints, delete orphans, reclaim space. - LCS unified diffs (
generateUnifiedDiff): O(n×m) longest common subsequence with 3-line context hunks. - cross-checkpoint search (
diffCheckpointsearch mode): Iterate checkpoints → resolve file hash → gunzip → line scan with ±2 context. - background snapshots (
worker.ts):spawn(detached)with.in-progresslockfile, <5ms return to Claude.
Disk usage: Vigil auto-skips derived directories (node_modules/, dist/, target/, venv/, etc.) detected from .gitignore and common patterns. Only source files are checkpointed.
| Project type | Raw size | Source only | First snapshot | Incremental |
|---|---|---|---|---|
| Next.js app | 750 MB | ~2 MB | ~1 MB | ~50 KB |
| Rust project | 2.5 GB | ~5 MB | ~3 MB | ~100 KB |
| Python project | 350 MB | ~3 MB | ~2 MB | ~50 KB |
After restore, vigil reports which derived dirs exist but weren't restored -- Claude rebuilds them (npm install, cargo build, etc.). Edit .claude/vigil/.vigilignore to adjust what gets skipped.
Architecture:
claude-vigil-mcp/
├── package.json
├── tsconfig.json
├── src/
│ ├── index.ts # MCP server, 5 tools
│ ├── types.ts # TypeScript interfaces and discriminated unions
│ ├── store.ts # CAS: hash, store, read, gc, disk usage
│ ├── snapshot.ts # create, restore, diff, list, delete
│ └── worker.ts # background snapshot process
├── .claude/
│ └── skills/
│ └── claude-vigil/
│ └── SKILL.md # optional skill for proactive checkpointing
└── test/
└── index.test.ts # 73 tests
Design principles:
- No git dependency -- pure Node.js built-ins (crypto, zlib, fs)
- Perfect snapshots -- every file captured, no size/binary filtering
- CAS + gzip -- 3.5x leaner than hard links, automatic dedup
- Background execution -- Claude never blocks on snapshot creation
- 3-slot limit -- conservative default prevents runaway storage
- Stateless server -- reads manifest from disk each call, no in-memory state to lose
- Artifact preservation -- displaced files saved on restore, nothing ever lost
- Cross-platform -- macOS, Linux, Windows. No shell dependencies
Design influences:
- Content-addressable storage -- same content = same address, automatic deduplication
- Mark-and-sweep garbage collection -- reference counting alternative for CAS cleanup
- Longest common subsequence -- foundation of unified diff generation
- Git object model -- inspiration for hash-sharded storage (vigil uses SHA-256 + gzip instead of git's SHA-1 + zlib)
alternatives
Every existing checkpoint tool -- built-in or third-party -- only tracks file edits made through the editor's own API. When Claude runs sed -i, a Python script overwrites a config, or a build tool corrupts output, those changes are invisible and unrecoverable.
| Feature | vigil | /rewind | Rewind-MCP | claude-code-rewind | Checkpoints app | Cursor |
|---|---|---|---|---|---|---|
| Tracks external changes | Yes (bash, scripts, builds) | No | No | No | No | No |
| Named checkpoints | Yes | No (timestamps) | Yes | Yes | Yes | No |
| Content diffs | Yes (unified diffs) | No | No | Yes (visual) | No | No |
| Search across checkpoints | Yes | No | No | No | No | No |
| Artifact preservation | Yes (nothing lost) | N/A | No | No | No | No |
| Dedup storage | CAS + gzip | None | Unknown | SQLite + diffs | Full copies | Zip per edit |
| Background saves | Yes (<5ms return) | Blocking | Blocking | Blocking | Background | Blocking |
| Headless/programmatic | Yes (MCP) | No (#16976) | Yes (MCP) | CLI | No | No |
| Cross-platform | Node.js | Built-in | Node.js | Python | macOS only | Built-in |
| Dependencies | 0 (Node built-ins) | N/A | Node.js | Python + SQLite | Desktop app | N/A |
| Disk visibility | Every response | Hidden | Manual | Manual | Manual | Hidden |
Claude Code /rewind -- Built-in checkpoint system. Only tracks file edits made through Claude's own tools -- sed -i, build scripts, and external processes are invisible. No named checkpoints, no content diffs, no search. Blocking saves that pause the session. Not yet available via MCP (#16976).
Rewind-MCP -- Third-party MCP server with stack-based undo. Provides MCP access but doesn't track external changes (bash, scripts, builds). No content diffs, no search across checkpoints, no dedup storage. Blocking saves.
claude-code-rewind -- Python-based snapshot tool with SQLite metadata and visual diffs. Requires Python + SQLite. Doesn't track external changes. No cross-checkpoint search, no artifact preservation, no background saves. CLI-only (no MCP integration).
Checkpoints app -- macOS desktop app that monitors projects for file changes. Background saves but macOS-only, no content diffs, no search, no MCP integration. Stores full copies of each checkpoint (no dedup), so disk usage grows linearly.
Cursor checkpoints -- Built-in to Cursor. Zips project state before each AI edit. No named checkpoints, no content diffs, no search, no external change tracking. Hidden from disk, blocking saves, not programmable.
Three implementation approaches were evaluated before settling on content-addressable storage:
- Shadow git repo (
git --git-dir=.claude/vigil/.git --work-tree=.) -- wraps git for dedup, diff, and restore. 6 failure modes: self-tracking recursion, unbounded binary bloat, concurrentindex.lockconflicts, overlay on restore (doesn't delete files added after checkpoint),git cleandestroying project files, andGIT_DIR/GIT_WORK_TREEenv var interference from parent processes. - Hard-link Time Machine pattern -- directory trees with unchanged files hard-linked (zero per-file cost). Battle-tested by macOS Time Machine, but 20 checkpoints of 1000 files = 20,000 directory entries. CAS + gzip is 3.5x leaner and deduplicates content across checkpoints for free.
- rsync --link-dest -- same hard-link idea in ~30 lines. No built-in diff capability, losing the unified diffs and cross-checkpoint search that make
vigil_diffuseful.
development
git clone https://github.com/Vvkmnn/claude-vigil-mcp && cd claude-vigil-mcp
npm install && npm run build
npm test
Scripts:
| Command | Description |
|---|---|
npm run build |
TypeScript compilation (tsc && chmod +x dist/index.js) |
npm run dev |
Watch mode (tsc --watch) |
npm start |
Run MCP server (node dist/index.js) |
npm run clean |
Remove build artifacts (rm -rf dist) |
npm run lint |
ESLint code quality checks |
npm run lint:fix |
Auto-fix linting issues |
npm run format |
Prettier formatting |
npm run format:check |
Check formatting without changes |
npm run typecheck |
TypeScript validation without emit |
npm test |
Type-check + lint |
npm run prepublishOnly |
Pre-publish validation (build + lint + format check) |
Pre-commit hooks via husky run lint-staged (prettier + eslint) on staged .ts files.
Contributing:
- Fork the repository and create feature branches
- Follow TypeScript strict mode and MCP protocol standards
Learn from examples:
- Official MCP servers for reference implementations
- TypeScript SDK for best practices
- Creating Node.js modules for npm package development
license
<hr>
<p align="center"><a href="https://en.wikipedia.org/wiki/Claudius#/media/File:Lebayle_-_Claudius.jpg"><img src="logo/lebayle-claudius.jpg" alt="Claudius Proclaimed Emperor -- Charles Lebayle" width="340"></a></p>
<p align="center">
Claudius Proclaimed Emperor by Charles Lebayle (1886). Claudius expanded the Vigiles Urbani from firefighters into Rome's night watch -- guardians who patrolled the city, preserved order, and ensured nothing was lost to the dark.
</p>
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。