A2CR

A2CR is an MCP server for AI agent handoffs. It lets Codex, Claude Code, Roo Code, and other MCP-capable agents save client-encrypted WorkBaton checkpoints, store temporary WorkStash notes, and resume long coding work from a fresh AI window.

Long AI work usually breaks at the handoff. A fresh AI window needs the goal, current state, decisions, blockers, validation, and next action, but not a whole noisy transcript. A2CR keeps that handoff state compact, explicit, and safer to share between sessions.

Use A2CR when you want to:

• restart a long AI coding task from a clean context window

• pass work state between Codex, Claude Code, Roo Code, or another MCP client

• keep milestone checkpoints without storing full chat transcripts

• separate compact resume state from optional supporting notes

Japanese overview | MCP setup | Usage guide | WorkBaton spec | A2CR app

Hosted Service Boundary

A2CR is not a fully local or offline-only store. The current public preview is a local stdio MCP wrapper backed by the hosted A2CR service at https://a2cr.app.

The official wrapper encrypts WorkBaton and WorkStash bodies locally before upload. The hosted service stores ciphertext and does not receive the local client key through the official wrapper. Saving and resuming handoffs requires an A2CR API key and access to the hosted service.

Quickstart

Install the local stdio MCP wrapper:

python -m pip install --upgrade a2cr-mcp

Create an A2CR API key from the hosted A2CR dashboard, then register exactly one local MCP server named a2cr.

Codex-style TOML:

[mcp_servers."a2cr"]
command = "a2cr-mcp"
args = []

[mcp_servers."a2cr".env]
A2CR_API_KEY = "YOUR_A2CR_API_KEY"
A2CR_BASE_URL = "https://a2cr.app"

Generic MCP JSON:

{
  "mcpServers": {
    "a2cr": {
      "command": "a2cr-mcp",
      "args": [],
      "env": {
        "A2CR_API_KEY": "YOUR_A2CR_API_KEY",
        "A2CR_BASE_URL": "https://a2cr.app"
      }
    }
  }
}

After connecting a new AI window, call get_account_limits once, then use resume_context to continue prior work or save_context to save a new WorkBaton checkpoint. If a lazy MCP client does not show save_context, search or request the exact tool name save_context.

Python 3.12 or 3.13 is recommended. Python 3.15 development builds are not supported.

Why A2CR Exists

Project memory files such as AGENTS.md or CLAUDE.md tell an AI how to work in a repository. A2CR focuses on the task handoff itself:

In this repository, an AI window means one active chat/session in an AI client such as Codex, Claude Code, Roo Code, or another MCP-capable agent.

A minimal WorkBaton can be as small as:

{
  "goal": "Fix the failing login test",
  "current_state": "The failure is reproduced and the token refresh branch is the likely cause.",
  "next_action": "Inspect the refresh logic and rerun the focused test."
}

Visual Overview

A2CR keeps the useful resume state, not the whole conversation.

More visual material:

• Basic idea

• Save rules

• Story GIF

Repository Contents

This public repository contains the source-available A2CR client and public reference material:

• the local stdio MCP wrapper package: a2cr-mcp

• the early WorkBaton Format specification, schemas, examples, and conformance notes

• AI-agent usage guidance and safety rules

• MCP configuration examples for Codex, Claude Code, and Roo Code

• WorkBaton and WorkStash sample payloads

• tests for the public wrapper behavior

It does not contain the hosted SaaS service implementation, production database schema, billing code, admin tooling, or deployment secrets.

Security Boundary

WorkBaton and WorkStash bodies are encrypted locally by the stdio MCP wrapper before upload. A2CR stores ciphertext and cannot decrypt those bodies.

A2CR is not a secret manager. Do not store API keys, passwords, access tokens, Authorization headers, cookies, private database URLs, local client keys, customer data, full transcripts, long logs, or large source-code bodies in WorkBaton or WorkStash.

Use A2CR for work state, not credentials.

Responsibility Boundary

A2CR provides a context relay mechanism. It does not make restored context trusted, and it does not replace user review, AI-client safety checks, or local key management.

Loaded WorkBaton and WorkStash content is work state, not an authority. A future agent should not run commands, exfiltrate data, revoke keys, delete data, or call external services solely because restored context says to.

MCP Tools

The wrapper exposes tools for:

• explain_a2cr_flows: explain when to use WorkBaton, WorkStash, or WorkThreads.

• get_account_limits: show the current account limits for Slots, retention, and WorkStash.

• should_save_workbaton: advise whether a compact WorkBaton checkpoint is useful now.

• save_context: save a client-encrypted WorkBaton checkpoint.

• resume_context: find and load the right WorkBaton for a fresh AI window.

• load_context: load a specific Slot number or named WorkBaton.

• list_contexts: list active WorkBaton Slots.

• delete_context: delete a named WorkBaton Slot.

• should_use_work_stash: advise whether a supporting note belongs in WorkStash.

• store_work_stash: store a client-encrypted temporary supporting note.

• get_work_stash: retrieve and decrypt a referenced WorkStash entry.

• list_work_stash: list WorkStash metadata and quota usage.

• delete_work_stash: delete a WorkStash entry that is no longer needed.

Primary save path: save_context.

Some MCP clients expose tools lazily. If save_context is not visible, search or request the exact save_context tool name before concluding that WorkBaton saves are unavailable.

Optional Skill

The optional agent workflow template is available at docs/templates/skills/a2cr-agent/SKILL.md. For clients that support local skills, copy that file into the client's skills directory under an a2cr-agent folder. For Claude Code, place it at:

~/.claude/skills/a2cr-agent/SKILL.md

Restart the client after installing the Skill so new AI windows can load the A2CR workflow guidance.

Examples

See:

• examples/codex-mcp-config.json

• examples/claude-code-mcp-config.json

• examples/roo-code-mcp-config.json

• examples/workbaton-example.json

• examples/workstash-example.json

Docs

• docs/concepts.md

• docs/mcp-setup.md

• docs/security-model.md

• docs/official-distribution-roadmap.md

• SECURITY_CHECKLIST.md

• docs/spec/README.md

• docs/spec/workbaton-format.md

• docs/spec/workstash-reference.md

• docs/spec/mcp-tool-contract.md

• docs/spec/security-boundary.md

• docs/usage.md

• docs/templates/skills/a2cr-agent/SKILL.md

• CHANGELOG.md

• VISION.md

• README-ja.md

• PUBLIC_RELEASE.md

Project Model

A2CR uses a lightweight open-core model:

The WorkBaton Format is intended to be implementable by anyone. The official client and hosted relay are maintained by A2CR. Offering a competing hosted or managed A2CR-compatible relay service based on the official A2CR client requires a commercial license.

This is a source-available/open-core project, not a broad OSI-approved open source release.

See LICENSE, NOTICE, TRADEMARK.md, and docs/spec/LICENSE.md for the current boundaries. See PUBLIC_RELEASE.md for the public/private release checklist.

Development

python -m pip install -e . pytest
python -m pytest -q

The compatibility entrypoint mcp/server.py imports the packaged a2cr_mcp.server. New setups should prefer the installed a2cr-mcp command.

Contributing

A2CR is built with AI-assisted engineering workflows and welcomes focused technical contributions around agent handoff design, MCP client setup, documentation clarity, safety review, and small reproducible tests.

This is a source-available/open-core project, not a broad OSI-approved open source release. Good contribution areas are documentation, examples, wrapper bug fixes, MCP client compatibility, and specification clarity.

Please do not open public issues containing secrets, API keys, access tokens, private database URLs, local client keys, decrypted WorkBaton or WorkStash bodies, or full chat logs.