# A2CR Manual
This page is the practical operating manual separate from the public concept guide and the AI agent guide.
From sign-in to first use:
- The current local stdio MCP wrapper is a Python program. The AI client starts it in the background, so Python is required.
- Recommended version: Python 3.13. Python 3.12 or newer is expected to work, but choose Python 3.13 if unsure.
- Avoid development builds such as Python 3.15 alpha or beta.
- Run python --version in a terminal to check the installed version.
- After Python is available, install the wrapper from PyPI with python -m pip install --upgrade a2cr-mcp.
- Sign in to A2CR and issue an API key from Settings.
- Do not paste the API key into AI chat. The user should paste it into the MCP config locally.
- In the MCP config, set command to a2cr-mcp and args to an empty array.
- Add exactly one MCP server named a2cr. Preserve existing MCP servers.
- In a new AI session, call get_account_limits and explain_a2cr_flows to verify the connection and learn the available flows.
Add to AGENTS.md, CLAUDE.md, or another project memory file:
- A2CR MCP tools may be used proactively when they help preserve useful work state.
- Use configured MCP tools only; do not invent direct HTTP API calls.
- Use WorkBaton at milestones, after validation, before likely context loss, or when context drift/contamination is detected.
- Use WorkStash for safe supporting notes that would bloat the WorkBaton.
- Record WorkStash entry_key values in WorkBaton references or next_action.
- Never save secrets, API keys, Authorization headers, cookies, private DB URLs, personal data, full transcripts, long logs, generated caches, git diffs, or large source-code bodies.
Saving to a Slot:
- Call get_account_limits to confirm plan limits and detail level.
- Call should_save_workbaton when the save is discretionary.
- Save compact goal, current_state, next_action, decisions, blockers, validation, and references with save_context.
- Move detailed file paths, API notes, failed attempts, and validation notes to WorkStash.
Reading from a Slot:
- In a fresh AI window, call resume_context(slot_number=N) or resume_context(slot_name="...").
- The AI reads goal, current_state, next_action, decisions, blockers, validation, and references.
- Slot metadata, response_language_hint, language_context, and agent_continuity_guidance may also be available.
- If WorkStash entry_key values are referenced, call get_work_stash only for entries needed for the current task.
What the AI receives from MCP:
- WorkBaton is a compact serial checkpoint handoff, not a chat log or file store.
- WorkStash is temporary supporting memory, not a durable knowledge base.
- The local stdio wrapper is the official WorkBaton save path. A2CR cannot decrypt client-encrypted WorkBaton bodies.
- The AI receives tool descriptions and safety rules for explain_a2cr_flows, get_account_limits, should_save_workbaton, save_context, resume_context, get_work_stash, and related tools.
Autonomy:
- With MCP tool descriptions, AGENTS.md / CLAUDE.md, and a loaded WorkBaton, capable AI agents can use WorkBaton and WorkStash proactively when needed.
- If context drift or context contamination appears, the agent can call should_save_workbaton, save a compact WorkBaton when recommended, and suggest continuing in a fresh AI window.