claude-checkpoint-mcp

Save your Claude Code session state locally, so a crash, /clear, or context limit doesn't make Claude rebuild context from scratch.

claude-checkpoint-mcp is a local Model Context Protocol server with three tools:

• save_session_checkpoint writes your session's working state to a small JSON file in the project.

• load_session_checkpoint turns that file back into a compact resume brief.

• list_session_checkpoints lists the checkpoints in the project, so you can pick which session to resume.

Each Claude Code session gets its own checkpoint, so several sessions can share a directory without overwriting each other.

It runs entirely on your machine. Nothing is sent anywhere.

Fork of tenlean/claude-checkpoint-mcp by Tenzin.

Contents

• Why: stop paying twice for the same context

• How it works

• Installation

• Configuration

  • Environment variables

  • Claude Code

  • Claude Desktop

  • Other MCP clients

• Usage

• Tool reference & examples

• The checkpoint files

• Security & privacy

• Troubleshooting

• Development

Related MCP server: handoff-protocol

Why: stop paying twice for the same context

A long Claude Code session builds up a lot of context:

• what you're trying to do;

• which approaches already failed;

• which files matter;

• which test is still red.

Several things can wipe that out mid-task:

• a terminal crash, SSH drop, or laptop sleep;

• running out of context window;

• a /clear to get a fresh window.

The next session starts cold. To get back to where you were, Claude has to rebuild that context:

• Re-reading files to work out what was changed and why. Every file read is input tokens.

• Re-running exploration: grep, git diff, ls, test runs, all just to rediscover state.

• Re-explaining. You retype the goal, the constraints, and the dead ends, or Claude rediscovers the dead ends the hard way.

Each of those turns resends the growing conversation as input, so the cost compounds.

claude-checkpoint-mcp swaps that rebuild for one small, structured read:

The resume brief for a typical session is about 1.2 KB of markdown, a few hundred tokens. The example below is real output at that size.

The biggest saving comes from drift detection. At save time, the server records a content hash for every file with uncommitted changes. On load, it compares those hashes against the working tree and tells Claude exactly which files changed after the checkpoint. Everything else in the brief can be trusted as-is, so Claude re-reads only what it has to.

How it works

  Session 1                                     Session 2 (after crash / /clear)
  ─────────                                     ────────────────────────────────
  Claude ──save_session_checkpoint──┐       ┌──load_session_checkpoint── Claude
         goal, tasks, summary,      │       │   compact markdown brief
         next steps, decisions,     ▼       │   + "changed since checkpoint"
         files, commands      ┌─────────────────────┐
                              │ <project>/          │
     server adds:             │ .claude-checkpoint  │◀── git: branch, HEAD,
     git branch/HEAD,         │ .claude-checkpoint  │    dirty files + hashes
     dirty files + hashes,    │   .prev (backup)    │
     redacts secrets          └─────────────────────┘

• Claude supplies the meaning: the goal, progress, decisions, next steps, and which commands mattered.

• The server supplies the facts: git branch and HEAD, every uncommitted or untracked file with its content hash, and timestamps.

• The server instructions teach Claude when to act. Claude is told to load at the start of a session and to save after milestones, before risky operations, and when the conversation gets long. It doesn't need to be prompted every time.

One checkpoint per session

Claude Code passes each MCP server its session id (CLAUDE_CODE_SESSION_ID). The server uses it to give every session its own file, <project>/.claude-checkpoints/<session-id>.json, so sessions that share a directory (/tmp, or wherever you run claude agents) never overwrite or merge into each other.

• Saving always writes the current session's own file.

• Loading returns the session's own checkpoint if it has one. Otherwise, if other checkpoints exist, it returns a list (goal, last-saved time, branch, whether all tasks are done) and Claude asks you which one to resume. It never picks one for you, even when there's only one.

• Resuming another session's checkpoint loads it, but saves still go to the current session's own file, starting from what was loaded and recording where it came from (resumedFrom). The original is never modified.

• Checkpoints are kept until you delete them, so an old session can be resumed later.

• Clients without a session id (e.g. Claude Desktop) use the single <project>/.claude-checkpoint file, exactly as before. A .claude-checkpoint left by an earlier version shows up in the list as legacy.

The session id an MCP server sees is the one it was started with:

Installation

Requirements: Node.js 20+ and git. Git is optional, but without it you lose drift detection.

git clone https://github.com/daniel-munoz/claude-checkpoint-mcp.git
cd claude-checkpoint-mcp
npm install
npm run build

The server entry point is now at dist/index.js.

From npm. The server is also published as session-checkpoint-mcp. In any of the configurations below, you can replace node /absolute/path/to/claude-checkpoint-mcp/dist/index.js with npx -y session-checkpoint-mcp:

claude mcp add --scope user checkpoint -- npx -y session-checkpoint-mcp

(The npm package claude-checkpoint-mcp is the upstream project's 0.1.0 release, not this fork.)

Configuration

Environment variables

Claude Code

As a plugin (easiest). This repo is also a Claude Code plugin marketplace. You don't need to clone or build anything:

claude plugin marketplace add daniel-munoz/claude-checkpoint-mcp
claude plugin install session-checkpoint@daniel-munoz

Restart Claude Code, then /mcp should show plugin:session-checkpoint:session-checkpoint as connected.

Upgrading from 0.2.x or earlier? The plugin was renamed from claude-checkpoint-mcp to session-checkpoint in 0.3.0, because Claude Code reserves the claude- prefix for its own plugins. Uninstall the old plugin and install the new one:

claude plugin uninstall claude-checkpoint-mcp@daniel-munoz
claude plugin marketplace update daniel-munoz
claude plugin install session-checkpoint@daniel-munoz

The full tool names changed from mcp__plugin_claude-checkpoint-mcp_claude-checkpoint__* to mcp__plugin_session-checkpoint_session-checkpoint__*. Update any permission allowlists in settings.json, hooks, or instructions that use the old names. Short names such as save_session_checkpoint and your existing checkpoint files are unaffected.

Manual setup. If you cloned the repo, register the server once for your user account, so it's available in every project:

# Run from the claude-checkpoint-mcp directory
claude mcp add --scope user checkpoint -- node "$PWD/dist/index.js"

Use an absolute path. A user-scoped server is launched from whichever project you open Claude Code in. A relative path like ./dist/index.js would resolve against that project, not this repo. "$PWD/dist/index.js" expands to the absolute path when you run the command, which is what gets saved.

Check that it's registered:

claude mcp list

Inside a session, /mcp should show checkpoint as connected.

Project-scoped alternative. To share the server with a team through version control, add it to the project's .mcp.json:

{
  "mcpServers": {
    "checkpoint": {
      "command": "node",
      "args": ["/absolute/path/to/claude-checkpoint-mcp/dist/index.js"]
    }
  }
}

Either way, add the checkpoint files to the project's .gitignore, or to your global gitignore:

.claude-checkpoint*

Claude Desktop (claude_desktop_config.json)

Open the config file:

Add the server under mcpServers, using the absolute path to your local build:

{
  "mcpServers": {
    "checkpoint": {
      "command": "node",
      "args": ["/absolute/path/to/claude-checkpoint-mcp/dist/index.js"],
      "env": {
        "CHECKPOINT_PROJECT_ROOT": "/absolute/path/to/your/project"
      }
    }
  }
}

Restart Claude Desktop after saving the file.

Claude Desktop has no notion of a "current project". Use CHECKPOINT_PROJECT_ROOT to set the default project, or have Claude pass projectRoot on each call ("save a checkpoint for /home/me/code/my-app"). See project root resolution.

Other MCP clients

Any MCP client that supports the stdio transport can run the server:

node /absolute/path/to/claude-checkpoint-mcp/dist/index.js
# or
npx -y session-checkpoint-mcp

Usage

The server's built-in instructions cover the common cases automatically. You can also ask Claude directly.

Save:

• "Checkpoint our progress."

• "Save a checkpoint before you run the migration."

• "Update the checkpoint. Merge in what we just did."

Resume (in a fresh session):

• "Resume where we left off."

• "Load the checkpoint and continue."

Example resume brief

This is real output from load_session_checkpoint. In the scenario, one file was committed unchanged after the save, and a teammate then edited session.ts:

# Resuming: my-app (checkpoint saved 2m ago, 2026-10-08 02:23:55 UTC)
Branch `feat/github-oauth` @ b2ef1b591 (uncommitted changes)

## Goal
Add GitHub OAuth login with persistent sessions

## Tasks
- [x] OAuth callback route
- [~] Session cookie persistence
- [ ] Logout endpoint

## Where things stand
Callback exchanges the code and creates a session, but the cookie is dropped on reload. Suspect SameSite/secure flags in dev.

## Next steps
1. Set sameSite: 'lax' and secure only in production in src/auth/session.ts
2. Re-run npm test -- auth
3. Start the logout endpoint

## Decisions & gotchas
- Use iron-session, not express-session (needs edge runtime support)
- Don't mock the GitHub API in e2e tests; use the recorded fixtures in test/fixtures

## Files touched (3)
- src/auth/session.ts (M) — cookie options
- src/auth/callback.ts (M)
- src/auth/session.spec.ts (??)

## Recent commands (last 2, oldest first)
- ✓ `GITHUB_CLIENT_SECRET=«redacted» npm run dev`
- ✗ exit 1 `npm test -- auth` — 2 failing in session.spec.ts: cookie missing after reload

## Since this checkpoint
- 1 new commit(s) on HEAD
- ⚠ Changed after the checkpoint, re-read before editing: `src/auth/session.ts`

callback.ts was committed after the save, but its content didn't change, so it isn't flagged. Only session.ts needs to be re-read.

Task markers: [x] done · [~] in progress · [ ] pending · [!] blocked.

Tool reference & examples

save_session_checkpoint

Writes the current session's checkpoint: <projectRoot>/.claude-checkpoints/<session-id>.json, or <projectRoot>/.claude-checkpoint when the client provides no session id.

Full save:

{
  "goal": "Add GitHub OAuth login with persistent sessions",
  "tasks": [
    { "description": "OAuth callback route", "status": "done" },
    { "description": "Session cookie persistence", "status": "in_progress" },
    { "description": "Logout endpoint", "status": "pending" }
  ],
  "summary": "Callback exchanges the code and creates a session, but the cookie is dropped on reload. Suspect SameSite/secure flags in dev.",
  "nextSteps": [
    "Set sameSite: 'lax' and secure only in production in src/auth/session.ts",
    "Re-run npm test -- auth"
  ],
  "decisions": ["Use iron-session, not express-session (needs edge runtime support)"],
  "modifiedFiles": [{ "path": "src/auth/session.ts", "note": "cookie options" }],
  "recentCommands": [
    { "command": "GITHUB_CLIENT_SECRET=gh_live_8f2a91 npm run dev", "exitCode": 0 },
    { "command": "npm test -- auth", "exitCode": 1, "note": "2 failing in session.spec.ts" }
  ]
}

Response (text shown to the model):

Checkpoint saved to /home/me/code/my-app/.claude-checkpoint (2202 bytes, replace).
3 task(s), 3 file(s), 2 command(s) on feat/github-oauth@b2ef1b591.
⚠ Redacted 1 secret-looking value(s).

Incremental update with merge:

{
  "goal": "Add GitHub OAuth login with persistent sessions",
  "mode": "merge",
  "tasks": [
    { "description": "Session cookie persistence", "status": "done" },
    { "description": "Logout endpoint", "status": "in_progress" }
  ],
  "recentCommands": [{ "command": "npm test -- auth", "exitCode": 0, "note": "all green" }]
}

Structured output (structuredContent):

{
  "path": "/home/me/code/my-app/.claude-checkpoint",
  "bytes": 2202,
  "mode": "replace",
  "counts": { "tasks": 3, "modifiedFiles": 3, "recentCommands": 2, "redactions": 1 },
  "warnings": ["Redacted 1 secret-looking value(s)."]
}

load_session_checkpoint

Read-only. Reads the checkpoint, checks for drift, and returns the brief.

Examples:

{}

{ "format": "full", "projectRoot": "/home/me/code/my-app" }

Structured output (structuredContent):

{
  "found": true,
  "path": "/home/me/code/my-app/.claude-checkpoint",
  "fromBackup": false,
  "updatedAt": "2026-10-08T02:23:55.477Z",
  "drift": {
    "branchChanged": false,
    "currentBranch": "feat/github-oauth",
    "commitsSince": 1,
    "changedSince": ["src/auth/session.ts"]
  },
  "warnings": []
}

If no checkpoint exists, the tool returns found: false with a "nothing to resume" message. That isn't treated as an error.

If the current session has no checkpoint but others exist, it returns found: false with candidates (the same entries as list_session_checkpoints, newest first, up to 10), total, omitted, and text telling Claude to ask you which one to resume. An unknown checkpoint selector is a tool error.

list_session_checkpoints

Read-only. Lists the project's checkpoints, newest first; unreadable ones are listed last with a warning rather than hidden.

Structured output (structuredContent, abridged):

{
  "path": "/home/me/scratch",
  "entries": [
    {
      "selector": "8ce2c093-d512-42c2-a4d8-aa98280b55df",
      "goal": "Investigate the flaky upload test",
      "updatedAt": "2026-10-19T21:40:23.000Z",
      "age": "2h ago",
      "branch": null,
      "allTasksDone": false,
      "isCurrentSession": false,
      "fromBackup": false
    },
    { "selector": "legacy", "goal": "Quick CSV analysis", "...": "..." }
  ],
  "total": 2,
  "omitted": 0,
  "warnings": [],
  "text": "2 checkpoint(s) in /home/me/scratch, newest first: …"
}

How the project root is chosen

1. The projectRoot argument

2. The CHECKPOINT_PROJECT_ROOT environment variable

3. The client's first file:// root, if the client supports MCP roots

4. The server process's working directory

The path must be absolute and must be an existing directory. Symlinks are resolved. Checkpoints are always written under exactly that root: <root>/.claude-checkpoints/<session-id>.json, or <root>/.claude-checkpoint without a session id.

The checkpoint files

<project>/
├── .claude-checkpoint              # legacy single checkpoint (clients without a session id, or 0.1.x)
├── .claude-checkpoint.prev         # its backup
└── .claude-checkpoints/            # one checkpoint per Claude Code session (directory mode 0700)
    ├── <session-id>.json
    └── <session-id>.json.prev      # each with its own backup

Every file is written atomically with mode 0600. The .claude-checkpoint* gitignore pattern covers all of them.

Each one is plain, versioned JSON, so it's easy to read, diff, or edit by hand. Real example:

{
  "version": 1,
  "id": "ee5507c0-855e-44a5-85c1-d97b767537c9",
  "createdAt": "2026-10-08T02:23:55.477Z",
  "updatedAt": "2026-10-08T02:23:55.477Z",
  "generator": { "name": "session-checkpoint-mcp", "version": "0.3.1" },
  "project": { "name": "my-app" },
  "git": {
    "branch": "feat/github-oauth",
    "head": "b2ef1b5919ee5d456ea146c9362a7abc92b1e556",
    "dirty": true
  },
  "goal": {
    "primary": "Add GitHub OAuth login with persistent sessions",
    "tasks": [
      { "description": "OAuth callback route", "status": "done" },
      { "description": "Session cookie persistence", "status": "in_progress" },
      { "description": "Logout endpoint", "status": "pending" }
    ]
  },
  "summary": "Callback exchanges the code and creates a session, but the cookie is dropped on reload. Suspect SameSite/secure flags in dev.",
  "nextSteps": [
    "Set sameSite: 'lax' and secure only in production in src/auth/session.ts",
    "Re-run npm test -- auth",
    "Start the logout endpoint"
  ],
  "decisions": [
    "Use iron-session, not express-session (needs edge runtime support)",
    "Don't mock the GitHub API in e2e tests; use the recorded fixtures in test/fixtures"
  ],
  "modifiedFiles": [
    { "path": "src/auth/session.ts", "status": "M", "source": "git", "note": "cookie options", "hash": "9ec893ef673dac860e7e2b6fcaad602983eeb0b1" },
    { "path": "src/auth/callback.ts", "status": "M", "source": "git", "hash": "f3ecc849c3875d4f11bb42ae0e5ab6c71931d7f1" },
    { "path": "src/auth/session.spec.ts", "status": "??", "source": "git", "hash": "9daeafb9864cf43055ae93beb0afd6c7d144bfa4" }
  ],
  "recentCommands": [
    { "command": "GITHUB_CLIENT_SECRET=«redacted» npm run dev", "exitCode": 0, "at": "2026-10-08T02:23:55.477Z" },
    { "command": "npm test -- auth", "exitCode": 1, "note": "2 failing in session.spec.ts: cookie missing after reload", "at": "2026-10-08T02:23:55.477Z" }
  ],
  "meta": { "redactions": 1, "truncated": false }
}

Paths are always relative to the project root. The file never stores absolute paths.

Security & privacy

• Local only. There are no network calls. The server reads git metadata and the project's checkpoint files, and writes only the current session's checkpoint.

• Secret redaction. The following are replaced with «redacted» in commands, notes, summaries, tasks and decisions before anything touches disk:

  • *_KEY=, *_TOKEN=, *_SECRET=, *PASSWORD= assignments, including Header: value and JSON forms

  • Authorization: Bearer … headers

  • credentials in URLs (postgres://user:pass@…)

  • --password / --token / --api-key / --secret flags

  • bare sk-…, ghp_…, github_pat_…, AKIA… and xox?-… tokens

  Redaction is a safety net, not a guarantee, so keep the file out of version control.

• Owner-only, crash-safe writes. Files are written with mode 0600. Each save writes a temp file, fsyncs it, and renames it into place, so a crash mid-save can't leave a half-written checkpoint.

• Automatic backup. The previous valid checkpoint is kept as .claude-checkpoint.prev. If the main file is corrupt, load falls back to the backup and says so. A corrupt file never overwrites a good backup.

• Bounded size. Every field has a length cap, and the file is capped at 64 KB (the oldest commands are dropped first), so a checkpoint can't grow into a token sink.

• No shell. Git runs through execFile with a 3-second timeout and GIT_OPTIONAL_LOCKS=0, so it never contends with your own git commands.

• Doesn't touch your config. If the checkpoint isn't git-ignored, the save response warns you. It never edits .gitignore for you.

• Safe with parallel tool calls. Tool calls are serialized inside the server, so concurrent merge saves can't lose updates.

Troubleshooting

Server logs go to stderr. Stdout is reserved for the MCP protocol.

Development

npm run typecheck   # strict TypeScript, src + tests
npm test            # unit tests + in-memory MCP end-to-end tests (vitest)
npm run build       # compile to dist/
npm run inspect     # open the MCP Inspector against dist/index.js

src/
├── index.ts    # entry point: stdio transport, signals, stderr logging
├── server.ts   # McpServer construction + server instructions
├── tools.ts    # save/load/list tool schemas and handlers
├── session.ts  # session id validation
├── listing.ts  # checkpoint listing for the "which one?" prompt
├── schema.ts   # zod schema for the checkpoint file (v1) + migrations
├── store.ts    # project root resolution, checkpoint paths, atomic I/O, backup/fallback
├── git.ts      # branch/HEAD/status snapshot, blob hashing, drift detection
├── redact.ts   # secret scrubbing
└── render.ts   # checkpoint → markdown resume brief

Releasing

The version lives in package.json, server.json (twice), plugin/.claude-plugin/plugin.json and .claude-plugin/marketplace.json. test/metadata.test.ts fails if they disagree. Bump them all, run npm run build, commit, then push a vX.Y.Z tag. The Publish workflow publishes session-checkpoint-mcp to npm and server.json to the MCP registry as io.github.daniel-munoz/claude-checkpoint-mcp.

If a release fails partway, don't re-run the failed job. Run the workflow again from Actions → Publish → Run workflow with the tag instead. Steps that already succeeded are skipped.

Built with the official @modelcontextprotocol/sdk and zod.

License

MIT