3Notch

3Notch

Local-first working-state checkpoints and portable handoff packets for AI coding agents. Continuation checkpoints recover Claude Code sessions after rate limits, crashes, and compaction; packets hand tasks across tools, repos, and machines — Markdown on disk, no cloud, no telemetry.

Category
Visit Server

README

3Notch

Save the working state your AI tools won't.

Website · Docs · Quickstart · npm

CI npm Node License: MIT 3notch MCP server

A Claude Code session dies mid-run — rate limit, crash, compaction, laptop sleep — or you're moving a task from Claude Code to Codex, Cursor, or ChatGPT. The code is still in git. The objective, decisions, and next steps often are not. The next session rebuilds that state from scratch.

3Notch is a local CLI and MCP server for:

  • Continuation checkpoints — when Claude Code is configured, recover after rate limits, model-down failures, and compaction
  • Portable packets — hand off selected work across tools, repos, and machines
  • Durable inbox — async delivery between stores that share a mailbox root

No cloud service. No account. No telemetry. Records stay on disk as Markdown (and optional artifacts) under .notch/.


Why this layer exists

Every major vendor ships transcript persistence and memory files. None ship durable working state:

  • Claude Code auto-saves transcripts and takes file checkpoints — but checkpoints are session-scoped undo, gone when the session ends.
  • OpenAI's Agents SDK serializes RunState — at planned human-approval pauses, not arbitrary failures.
  • Gemini and Grok persist conversations server-side. A transcript is a diary, not a manifest.

Recovering from a transcript means re-reading a conversation and re-interpreting what happened. The vendors' own guidance converges on the fix: combine native persistence with handoff files and checkpoint strategies. 3Notch is that layer — local, vendor-neutral, yours.


Install

npm install -g @3notch/cli
notch onboard

Agent prompt (optional):

Install @3notch/cli, run notch onboard in this repo, and set up continuation checkpoints if I use Claude Code. Use packets for handoffs across tools or repos. Read the package README before changing MCP config.


Common flows

Resume after a failed session

With continuation checkpoints enabled for Claude Code, 3Notch can write a fallback from task state and git when the session hits a rate limit, model-down failure, or compaction. The next session may offer that checkpoint once; you approve before it is loaded.

notch packet list
notch packet preview <id>

Hand off between tools

notch packet create \
  --title "Auth refactor checkpoint" \
  --summary "Token validation done; session store migration blocked." \
  --next-steps "Implement Redis session adapter"

notch packet preview <id>
# other tool / session imports and continues

Ship files to another repo or machine

notch packet create \
  --title "Brand handoff" \
  --summary "Assets and layout for the launch page." \
  --file mascot.jpg:asset \
  --file showcase.html:source \
  --next-steps "Build the launch page from showcase.html and mascot.jpg."

notch packet pack <id>
# move <id>.notchpkt however you prefer, then:
notch packet unpack <id>.notchpkt

Async agents — durable inbox

Both sides register the same mailbox root, then pack/send and pull/ack:

notch inbox init --name review-agent --root /shared/3notch-mailbox
notch packet pack <id>
notch send <id>.notchpkt --to local:review-agent
# forward the printed delivery notice via chat, Slack, etc.

notch inbox list
notch inbox pull <delivery-id> --import
notch inbox ack <delivery-id>

local: addresses are routing labels, not authenticated identity. See Durable inbox.

Web chat without MCP

notch prompt --client claude-chat
# paste into the chat, copy the packet back
pbpaste | notch packet import -

Personal capture

notch mark --summary "Keep browser auth cookie-based" --tags auth

How it works

  1. You or an agent write selected context through the CLI or MCP tools.
  2. 3Notch validates, secret-scans, and stores records under .notch/.
  3. Preview before another agent relies on the content.
  4. The next session, tool, or store imports or resumes from that record.

Targeting fields (--to-agent, --to-repo) are intent metadata. Bytes move via your transport (scp, git, AirDrop, Tailscale) or the durable inbox mailbox — not a 3Notch-hosted relay.


Commands

notch onboard                         initialize .notch/ and MCP setup
notch packet create                   create a packet (--file, --ref, --next-steps)
notch packet import <path>            import into .notch/inbox/
notch packet preview <id>             show what an agent will read
notch packet pack / unpack            .notchpkt archive round-trip
notch packet list / show              list / inspect packets
notch inbox init/list/status/pull/ack durable delivery lifecycle
notch send <archive> --to <address>   send a packed project handoff
notch reply <id>                      typed reply to a packet
notch mark                            self-addressed private capture
notch brief / brief create|list|show  scoped task briefs
notch seed from <path>                private context seeding
notch prompt --client <client>        agent / web-chat instruction packs
notch scan <file-or-stdin>            secret scanner
notch check                           structural corpus checks
notch doctor                          store diagnostics
notch status                          store summary
notch mcp serve                        local stdio MCP server

MCP

notch mcp serve over local stdio:

Tools
Read get_brief, list_briefs, get_targeted_brief, get_packet, list_packets, list_inbox, get_inbox_delivery, get_status, check_store, run_doctor
Write create_brief, create_packet, create_mark, create_reply, create_seed_packet, import_packet, import_seed_packet, inbox_init, send_packet, pull_inbox_packet, ack_inbox_delivery

Private records under .notch/private/ stay hidden unless the server starts with --include-private. Client setup: docs/guides/mcp-setup.md.


Documentation

Topic Guide
Index docs/README.md
Cross-repo packets docs/guides/cross-repo-packets.md
Cross-tool handoff docs/guides/cross-tool-handoff.md
Durable inbox docs/guides/durable-inbox.md
Continuation checkpoints docs/guides/continuation-checkpoints.md
MCP setup docs/guides/mcp-setup.md
Privacy docs/reference/privacy.md
Security docs/reference/security-story.md
Releases (maintainers) docs/guides/release.md

Website docs mirror: https://3notch.dev/docs/.


Boundaries

  • Local files by default — no hosted relay, account system, or telemetry
  • No vector DB / native DB dependency
  • No arbitrary shell execution through MCP
  • You move bytes; 3Notch validates, scans, hashes, and stores them

Regression guard: tests/unit/no-deferred-commands.test.ts.


Contributing

Prefer opening an issue before large features. See open issues and CONTRIBUTING.md.

git clone https://github.com/coldlogicAI/3notch.git
cd 3notch
npm install
npm run lint && npm run type-check && npm run build
npm test && npm run test:e2e
npm run release:check

See CONTRIBUTING.md.

Architecture reference (historical but still useful): docs/archived-plans/v1/3notch-v1-technical-spec.md.


License

MIT © 3Notch contributors

Recommended Servers

playwright-mcp

playwright-mcp

A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.

Official
Featured
TypeScript
Magic Component Platform (MCP)

Magic Component Platform (MCP)

An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.

Official
Featured
Local
TypeScript
Audiense Insights MCP Server

Audiense Insights MCP Server

Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.

Official
Featured
Local
TypeScript
VeyraX MCP

VeyraX MCP

Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.

Official
Featured
Local
graphlit-mcp-server

graphlit-mcp-server

The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.

Official
Featured
TypeScript
Kagi MCP Server

Kagi MCP Server

An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.

Official
Featured
Python
Neon Database

Neon Database

MCP server for interacting with Neon Management API and databases

Official
Featured
E2B

E2B

Using MCP to run code via e2b.

Official
Featured
Exa Search

Exa Search

A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.

Official
Featured
Qdrant Server

Qdrant Server

This repository is an example of how to create a MCP server for Qdrant, a vector search engine.

Official
Featured