Agents can set up recipes, start batches, read logs and failure analyses, and retry. Give your agent two things:
bakery CLI instead.list_projects list_batches start_batch get_batch get_job_log explain_failure cancel_batch validate_recipe list_ovens list_deliveries deliver_batch update_delivery test_delivery list_accounts test_account list_ingredients get_screenshots
https://buildbakery.com/api/mcp. No token needed.The same two pieces, per agent. Pick yours:
Commit the files so teammates' agents and cloud agents get them too.
curl -fsSL https://buildbakery.com/install-cli.sh | sh bakery ai install --tool claude
For every project on this machine instead: bakery ai install --tool claude --global
Writes .claude/skills/buildbakery/SKILL.md. To install by hand, save the file at that path.
Claude.ai, desktop and mobile apps: Settings → Connectors → Add custom connector, URL https://buildbakery.com/api/mcp, then Connect, sign in to BuildBakery and Allow (all projects or one). Upload buildbakery-skill.zip under Settings → Capabilities → Skills so Claude also knows how to use it. Headless or CI instead of a browser sign-in: add --header "Authorization: Bearer $BAKERY_TOKEN" to the claude mcp add command.
Agents use the MCP tools when connected and fall back to the CLI. Agents that support it sign in through your browser and you approve them (all projects or one); otherwise use an API token from Account → API tokens. MCP setup for Claude (Claude Code, all projects; then /mcp → buildbakery → Authenticate):
claude mcp add --scope user --transport http buildbakery https://buildbakery.com/api/mcp claude mcp list # buildbakery should show as connected after /mcp → Authenticate
With pull request builds turned on, each change an agent pushes gets a PR comment with an install link, which works when you're driving the agent from a phone.