# A2CR Guide
A2CR saves AI work state as a WorkBaton so another AI window, model, or MCP-capable client can resume the work.
A2CR is not an AI. It is the baton that lets AI agents hand work to one another.
Let your AI read it:
- Show the AI agent guide to the AI agent you already use and ask it to explain A2CR.
- The guide is written for agents, so it can turn the app's role, limits, and setup into plain guidance for your situation.
Important points:
- Install the local stdio MCP wrapper with python -m pip install --upgrade a2cr-mcp.
- In MCP config, set command to a2cr-mcp, args to an empty array, and MCP server name to a2cr.
- A2CR is listed in the official MCP Registry as io.github.a2cr/a2cr-mcp.
- MCP Registry: https://registry.modelcontextprotocol.io/v0.1/servers/io.github.a2cr%2Fa2cr-mcp/versions/latest
- GitHub Release v0.1.6: https://github.com/a2cr/a2cr/releases/tag/v0.1.6
- Claude Desktop MCPB: https://github.com/a2cr/a2cr/releases/download/v0.1.6/a2cr-0.1.6.mcpb
- PyPI: https://pypi.org/project/a2cr-mcp/0.1.6/
- WorkBaton is client-encrypted only.
- A2CR does not accept plaintext WorkBaton bodies.
- The local stdio MCP wrapper encrypts before upload, and A2CR stores and returns ciphertext only.
- Direct remote HTTP MCP saving is disabled for WorkBaton.
- The local client key is managed by the user, not by the service administrator.
- The full API key is shown only once when issued. Reissuing creates a different API key, so MCP configs must be updated.
- The local stdio MCP wrapper creates the local client key file during the first client-encrypted save when no key file exists.
- Set A2CR_CLIENT_KEY_FILE to choose the exact key file path, or A2CR_CONFIG_DIR to choose the directory that contains workbaton.key.
- If neither variable is set, the default path is %APPDATA%\A2CR\workbaton.key on Windows, and $XDG_CONFIG_HOME/a2cr/workbaton.key or ~/.config/a2cr/workbaton.key on macOS/Linux.
- To resume from another PC, the user needs the A2CR API key plus the same local client key file. The API key alone cannot decrypt saved WorkBaton bodies.
- If the local client key is lost, old client-encrypted slots cannot be recovered by A2CR.
- Slots saved after creating a new local client key can be read with that new key, but old slots still need the old key.
Infrastructure and data boundaries:
- Hosted A2CR uses Supabase/Postgres for the data layer and Railway for the app runtime.
- User-owned rows are isolated with Supabase Row Level Security (RLS) and the least-privileged a2cr_app runtime role.
- Official WorkBaton saves are encrypted locally before upload, so A2CR stores ciphertext for WorkBaton bodies.
- Supabase and Railway publish SOC 2 / compliance information for their platforms. That helps with vendor risk, but it does not make A2CR itself SOC 2 certified and does not replace A2CR's own RLS, client encryption, key hygiene, and smoke tests.
Related pages:
- AI agent guide: https://a2cr.app/en/agent-guide
- Japanese guide: https://a2cr.app/guide
- MCP service URL: https://a2cr.app/mcp
Compression / summarization vs A2CR / WorkBaton:
- Compression and summarization shorten a long conversation.
- A2CR / WorkBaton passes a state that lets the next AI resume work.
- Compression and summarization target history inside that chat.
- A2CR / WorkBaton targets another chat, another AI, or another tool.
- Compression and summarization output summary text.
- A2CR / WorkBaton outputs work state such as goal, current_state, next_action, and blockers.
- Compression is a diet for a conversation log. A2CR is a baton for work state.
Sub-agents vs A2CR:
- Sub-agents divide work inside the same environment.
- A2CR carries the environment itself forward.
- Sub-agents are effective inside that chat and its parent agent.
- A2CR can work across ChatGPT, Claude, Codex, Cursor, and local LLMs.
- Sub-agent state sharing depends on the parent agent's context.
- A2CR stores state in an external temporary relay DB and expires it by TTL.
MCP / A2A / A2CR:
- MCP connects an AI agent to tools, APIs, and external data.
- A2A connects AI agents to other AI agents for delegation, communication, and collaboration.
- A2CR preserves compact work state across AI windows, models, tools, and time so the next session can resume from a clean handoff.
- A2CR is complementary to MCP and A2A, not a replacement for either protocol.