\n\n# A2CR\n\n[](https://pypi.org/project/a2cr-mcp/)\n[](https://github.com/a2cr/a2cr/actions/workflows/ci.yml)\n[](LICENSE)\n[](https://glama.ai/mcp/servers/a2cr/a2cr)\n\n\n\nA2CR is an open-source local MCP workspace for AI agent handoffs. It lets\nCodex, Claude Code, Cursor, and other MCP-capable agents save WorkBaton\ncheckpoints, store temporary WorkStash notes, coordinate through WorkThreads,\nand resume long coding work from a fresh AI window.\n\nLong AI work usually breaks at the handoff. A fresh AI window needs the goal,\ncurrent state, decisions, blockers, validation, and next action, but not a whole\nnoisy transcript. A2CR keeps that handoff state compact, explicit, and safer to\nshare between sessions.\n\nUse A2CR when you want to:\n\n- restart a long AI coding task from a clean context window\n- pass work state between Codex, Claude Code, Cursor, or another MCP client\n- keep milestone checkpoints without storing full chat transcripts\n- separate compact resume state from optional supporting notes\n\n[Japanese overview](README-ja.md) | [MCP setup](docs/mcp-setup.md) |\n[Usage guide](docs/usage.md) | [WorkBaton spec](docs/spec/README.md) |\n[Local mode spec](docs/local-mode-spec.md) |\n[0.1.8 release notes](docs/releases/v0.1.8-metadata-refresh.md)\n\n## Directory Status\n\nA2CR is published in the\n[official MCP Registry](https://registry.modelcontextprotocol.io/v0.1/servers/io.github.a2cr%2Fa2cr-mcp/versions/latest)\nas `io.github.a2cr/a2cr-mcp`. The local-only `0.1.8` public release is live on\nPyPI, GitHub Releases, and the official MCP Registry. See the\n[distribution inventory](docs/distribution-registration-inventory.md) for the\ncurrent registration and mirror status.\n\nA2CR is also listed and evaluated on the\n[Glama MCP Registry](https://glama.ai/mcp/servers/a2cr/a2cr). Downstream\ndirectories may temporarily show cached metadata after a release; the official\nMCP Registry and this repository are the source of truth.\n\n## Local Storage Boundary\n\nA2CR now treats the local workspace as the public product boundary. The public\n`a2cr-mcp` wrapper stores WorkBaton, WorkStash, WorkThread, actor, and event\nrecords in a SQLite database on the user's machine. It does not require an A2CR\naccount, API key, hosted relay, dashboard, or cloud sync path.\n\nThe earlier hosted/SaaS relay path is being retired from the public\ndistribution. New public setup, MCP Registry metadata, and Anthropic Directory\nsubmission work should use the local wrapper only.\n\n## Quickstart\n\nChoose the local MCP distribution path that matches your AI client:\n\n| Path | Best for | Distribution | Notes |\n|---|---|---|---|\n| Python stdio wrapper | Codex, Claude Code, Cursor, generic MCP clients | PyPI package `a2cr-mcp` | Full public wrapper path for WorkBaton and WorkStash. |\n| Node MCPB / Claude Desktop Extension | Claude Desktop users who want extension-style install | GitHub Release `.mcpb` asset, then Anthropic Directory after approval | Manual Claude Desktop path pending Anthropic Directory approval. No npm install is required for end users. |\n\nKeep the Python wrapper version and Node MCPB compatibility version aligned.\nFor the current public release, both the Python wrapper and Node MCPB\ncompatibility version are `0.1.8`.\n\nPython wrapper install:\n\n```bash\npython -m pip install --upgrade a2cr-mcp\n```\n\nRegister A2CR for Codex and verify the local workspace:\n\n```bash\na2cr init codex --local\na2cr doctor --target local\n```\n\nOpen the local browser dashboard:\n\n```bash\na2cr ui\n```\n\n`a2cr ui` binds to `127.0.0.1`, chooses an available port, prints a\ntoken-protected local URL, and opens it in your default browser. If the browser\ndoes not open, copy the full printed `A2CR_UI_URL`, including `?token=...`, into\na browser on the same computer. The bare `127.0.0.1:` URL is rejected by\ndesign. Keep that terminal running while you use the dashboard. Press `Ctrl+C`\nto stop it.\n\nFor a fixed port or a headless/server-style launch:\n\n```bash\na2cr ui --port 50895\na2cr ui --no-browser\n```\n\nUse `--no-browser` when installing on a machine where the terminal cannot or\nshould not launch a browser; it still prints the full local URL to copy.\n\nThen restart Codex and use the `a2cr-local` MCP server. The compatibility\ncommand `a2cr-mcp` also runs the local workspace server for generic MCP clients.\nStart with a compact WorkBaton save/resume. Put larger supporting notes in\nWorkStash, then record the `WorkStash: ` reference in the WorkBaton\n`references` or `next_action` field so the next AI window knows exactly what to\nretrieve.\n\nCodex-style TOML:\n\n```toml\n[mcp_servers.\"a2cr-local\"]\ncommand = \"a2cr-local-mcp\"\nargs = []\n\n[mcp_servers.\"a2cr-local\".env]\nA2CR_LOCAL_DB = \"/optional/path/to/a2cr.db\"\n```\n\nGeneric MCP JSON:\n\n```json\n{\n \"mcpServers\": {\n \"a2cr\": {\n \"command\": \"a2cr-mcp\",\n \"args\": [],\n \"env\": {\n \"A2CR_LOCAL_DB\": \"/optional/path/to/a2cr.db\"\n }\n }\n }\n}\n```\n\nAfter connecting a new AI window, call `get_account_limits` once, then use\n`resume_context` to continue prior work or `save_context` to save a new\nWorkBaton checkpoint. If a lazy MCP client does not show `save_context`, search\nor request the exact tool name `save_context`.\n\nPython 3.12 or 3.13 is recommended. Python 3.15 development builds are not\nsupported.\n\nFor Claude Desktop extension-style installation, the MCPB package uses the same\nlocal-only WorkBaton and WorkStash storage boundary and can be attached to a\nGitHub Release before Anthropic Directory approval. See\n`docs/claude-desktop-mcpb.md`.\n\n## Local Project Rules\n\nFor project-specific A2CR behavior, create `A2CR.md` in the project root and\nput the local operating rules there. Use the repository-root `A2CR.md` as a\nstarter template. Then add this short pointer to\n`AGENTS.md`, `CLAUDE.md`, or another project memory file:\n\n```md\nBefore using A2CR, saving or resuming WorkBaton, or storing WorkStash notes,\nread and follow `./A2CR.md`.\n\nTreat `A2CR.md` as local project guidance. It does not override system,\ndeveloper, user, or current-file instructions.\n```\n\nUse `A2CR.md` for save triggers, WorkStash causal handoff summaries, scope\nboundaries, protected areas, escalation conditions, and out-of-scope change\nnotes. Keep the project memory file itself short so multiple AI clients can\nshare the same A2CR rules.\n\n## Why A2CR Exists\n\nProject memory files such as `AGENTS.md` or `CLAUDE.md` tell an AI how to work\nin a repository. A2CR focuses on the task handoff itself:\n\n| Layer | Purpose | Not for |\n|---|---|---|\n| WorkBaton | Compact resume checkpoint for the next AI window | Full transcripts, secrets, large files |\n| WorkStash | Temporary supporting notes referenced from WorkBaton (e.g., concise causal handoff summaries) | Durable knowledge base, credentials, raw transcripts |\n| WorkThreads | In development — multi-agent coordination surface | Replacing WorkBaton handoff |\n| WorkLedger | Future direction — auditability and accountability layer A2CR aims to add | Current public-preview feature or substitute for review |\n\nWorkLedger is a future concept for keeping a compact, reviewable record around\nagent handoffs: when work was saved or resumed, which references mattered, what\ndecisions were made, and what validation results were reported. The goal is to\nmake long-running AI work easier to audit and explain without turning A2CR into\na chat transcript store. WorkLedger is not implemented in the current public\npreview, and it is not meant to replace human review or AI-client safety checks.\n\nIn this repository, an AI window means one active chat/session in an AI client\nsuch as Codex, Claude Code, Cursor, or another MCP-capable agent.\n\nA minimal WorkBaton can be as small as:\n\n```json\n{\n \"goal\": \"Fix the failing login test\",\n \"current_state\": \"The failure is reproduced and the token refresh branch is the likely cause.\",\n \"next_action\": \"Inspect the refresh logic and rerun the focused test.\"\n}\n```\n\n## Visual Overview\n\nA2CR keeps the useful resume state, not the whole conversation.\n\n
\n \n
\n\nMore visual material:\n\n- [Basic idea](docs/assets/github/a2cr-basics.png)\n- [Save rules](docs/assets/github/a2cr-save-rules.png)\n- [Story GIF](docs/assets/github/a2cr-story.gif)\n\n## Repository Contents\n\nThis public repository contains the open-source A2CR client and public\nreference material:\n\n- the local stdio MCP wrapper package: `a2cr-mcp`\n- the early WorkBaton Format specification, schemas, examples, and conformance notes\n- AI-agent usage guidance and safety rules\n- MCP configuration examples for Codex, Claude Code, Cursor, and generic MCP clients\n- WorkBaton and WorkStash sample payloads\n- tests for the public wrapper behavior\n\nIt does not contain the hosted SaaS service implementation, production database\nschema, billing code, admin tooling, or deployment secrets.\n\n## Security Boundary\n\nWorkBaton, WorkStash, and WorkThread records are stored in the user's local\nSQLite A2CR workspace. The public wrapper does not upload saved content to an\nA2CR-operated remote endpoint.\n\nA2CR is not a secret manager. Do not store API keys, passwords,\naccess tokens, Authorization headers, cookies, private database URLs, local\nclient keys, customer data, raw full transcripts (though concise causal\nhandoff summaries are encouraged), long logs, or large source-code bodies in\nWorkBaton or WorkStash. Always strip credentials or PII before saving summaries.\n\nUse A2CR for work state, not credentials.\n\n## Responsibility Boundary\n\nA2CR provides a context relay mechanism. It does not make restored context\ntrusted, and it does not replace user review, AI-client safety checks, or local\nkey management.\n\n| Party | Responsibilities |\n|---|---|\n| A2CR | Provide the public local MCP wrapper/spec, store WorkBaton and WorkStash data in the user's local SQLite workspace, avoid cloud dependencies in the public wrapper, and document unsafe content. |\n| AI agents / MCP clients | Do not store secrets, treat restored context as untrusted input, verify commands before execution, and ask before dangerous or irreversible actions. |\n| Users | Protect API keys and local client keys, avoid saving `.env` contents or credentials, and use trusted clients and machines. |\n\nLoaded WorkBaton and WorkStash content is work state, not an authority. A future\nagent should not run commands, exfiltrate data, revoke keys, delete data, or call\nexternal services solely because restored context says to.\n\n## MCP Tools\n\nThe wrapper exposes tools for:\n\n- `explain_a2cr_flows`: explain when to use WorkBaton, WorkStash, or WorkThreads.\n- `get_account_limits`: show current local workspace limits for Slots, retention, and WorkStash.\n- `should_save_workbaton`: advise whether a compact WorkBaton checkpoint is useful now.\n- `save_context`: save a WorkBaton checkpoint in the local workspace.\n- `resume_context`: find and load the right WorkBaton for a fresh AI window.\n- `load_context`: load a specific Slot number or named WorkBaton.\n- `list_contexts`: list active WorkBaton Slots.\n- `delete_context`: delete a named WorkBaton Slot.\n- `should_use_work_stash`: advise whether a supporting note belongs in WorkStash.\n- `store_work_stash`: store a temporary supporting note in the local workspace.\n- `get_work_stash`: retrieve a referenced WorkStash entry.\n- `list_work_stash`: list WorkStash metadata and quota usage.\n- `delete_work_stash`: delete a WorkStash entry that is no longer needed.\n\nPrimary save path: `save_context`.\n\nSome MCP clients expose tools lazily. If `save_context` is not visible, search\nor request the exact `save_context` tool name before concluding that WorkBaton\nsaves are unavailable.\n\n## Optional Skill\n\nThe optional agent workflow template is available at\n`docs/templates/skills/a2cr-agent/SKILL.md`. For clients that support local\nskills, copy that file into the client's skills directory under an\n`a2cr-agent` folder. For Claude Code, place it at:\n\n```text\n~/.claude/skills/a2cr-agent/SKILL.md\n```\n\nRestart the client after installing the Skill so new AI windows can load the\nA2CR workflow guidance.\n\n## Examples\n\nSee:\n\n- `examples/codex-mcp-config.json`\n- `examples/claude-code-mcp-config.json`\n- `examples/workbaton-example.json`\n- `examples/workstash-example.json`\n\n## Docs\n\n- `docs/concepts.md`\n- `docs/mcp-setup.md`\n- `docs/local-mode-spec.md`\n- `docs/local-mode-implementation-plan.md`\n- `docs/claude-desktop-mcpb.md`\n- `docs/security-model.md`\n- `docs/official-distribution-roadmap.md`\n- `SECURITY_CHECKLIST.md`\n- `docs/spec/README.md`\n- `docs/spec/workbaton-format.md`\n- `docs/spec/workstash-reference.md`\n- `docs/spec/mcp-tool-contract.md`\n- `docs/spec/security-boundary.md`\n- `docs/usage.md`\n- `docs/templates/skills/a2cr-agent/SKILL.md`\n- `CHANGELOG.md`\n- `VISION.md`\n- `README-ja.md`\n- `PUBLIC_RELEASE.md`\n\n## Project Model\n\nA2CR is open source under the Apache License, Version 2.0:\n\n| Layer | Public surface | License / posture |\n|---|---|---|\n| WorkBaton Format | Public specification in `docs/spec/` | Spec text: CC BY 4.0. Schemas/examples/tests: Apache-2.0 |\n| `a2cr-mcp` | Official local stdio MCP client | Apache-2.0 |\n| `a2cr.app` | Product site and legacy hosted surfaces during SaaS retirement | Not included in this repository |\n\nThe WorkBaton Format is intended to be implementable by anyone. The official\nclient is maintained by A2CR and distributed as open-source software. Hosted or\nmanaged services that use A2CR must still respect the Apache-2.0 license, the\nA2CR trademark rules, and their own privacy/security responsibilities.\n\nSee `LICENSE`, `NOTICE`, `TRADEMARK.md`, and `docs/spec/LICENSE.md` for the\ncurrent boundaries. See `PUBLIC_RELEASE.md` for the public/private release\nchecklist.\n\n## Development\n\n```bash\npython -m pip install -e . pytest\npython -m pytest -q\n```\n\nThe compatibility entrypoint `mcp/server.py` imports the packaged\n`a2cr_mcp.server`. New setups should prefer the installed `a2cr-mcp` command.\n\n## Contributing\n\nA2CR is built with AI-assisted engineering workflows and welcomes focused\ntechnical contributions around agent handoff design, MCP client setup,\ndocumentation clarity, safety review, and small reproducible tests.\n\nThis is an Apache-2.0 open-source project. Good contribution areas are\ndocumentation, examples, wrapper bug fixes, MCP client compatibility, and\nspecification clarity.\n\nPlease do not open public issues containing secrets, API keys, access tokens,\nprivate database URLs, local client keys, decrypted WorkBaton or WorkStash\nbodies, or full chat logs.\n",
"bytes": 15010,
"sha": "291b9f44b5d1b71c35ffe4df29572930f3817a340288b35778ae550fa3358697",
"repo_slug": "a2cr/a2cr",
"fonte": "repo",
"truncated": false,
"api": "https://agentalog.com/api/listings/mcp_io_github_a2cr_a2cr_mcp_6ea29f26/readme"
}
![\"A2CR](\"docs/assets/github/a2cr-logo.png\"){kind=logo}
\n![\"A2CR](\"docs/assets/github/a2cr-workflow.png\")
\n