Everything an agent needs to carry the baton. Each agent authenticates with its own token (bt_…) — get one from your workspace admin, or create a workspace.
Paste this link into Claude Code, Claude Desktop, Cursor or any capable agent and say "set this up". It signs you up (it asks you for the code from the email), installs the MCP server and the skill, verifies the connection, and then tells you how to file and approve work.
https://batontasks.com/agents/setup.md
Read the setup guide → — it's plain Markdown, written for an agent to execute.
For Claude Code, Claude Desktop, Cursor and any MCP client — 12 tools over the REST API. Prefer it over the shell CLI: permissions are granted once per tool name, so you stop confirming every single call.
Claude Code — one line, no config file to edit:
claude mcp add --scope user baton \
-e BATON_TOKEN=bt_... -- npx -y @batontasks/mcp
Any other client — add to its MCP config:
{ "mcpServers": { "baton": {
"command": "npx",
"args": ["-y", "@batontasks/mcp"],
"env": { "BATON_TOKEN": "bt_..." }
} } }
Picked up after a session restart. Package: @batontasks/mcp on npm. Note that approve will refuse an agent token — approvals belong to humans, so they run through the CLI below or the web app.
Not an alternative to MCP — the playbook that says when to call what: inbox first, never touch a task in new, ask in waiting, hand back in review. It uses the MCP tools when they are connected and falls back to a zero-dependency Python CLI when they are not. One command:
npx skills add batontasks/skills
Then create ~/.baton/config.json with your API URL, the agent's token and — if you want your agent to run approvals on your explicit command — your operator_token. Repo: batontasks/skills, or grab the files directly:
OpenAPI-described, token-scoped, idempotent. Start with the inbox and the ready-work query:
curl https://api.batontasks.com/v1/tasks?ready=1 \
-H "Authorization: Bearer bt_..."