Docs · Claude Code plugin
Plan with Corvi in Claude Code.
One command, and your Corvi plan library is in Claude Code — read a plan, fan its todos into parallel-safe waves, and check a diff against it. It runs on the Corvi app already on your Mac: reads stay local, and your agent asks before any tool runs.
Install
The plugin needs the Corvi app — it runs against the local server the app ships. In Claude Code, add the marketplace, then install:
/plugin marketplace add corvi-sh/claude-plugin
/plugin install corvi@corviIf Corvi isn't in /Applications, point the plugin's corvi_binary config at Corvi.app/Contents/Resources/corvi.
Use it
Type /corvi [query] to start from a matching plan, or just ask your agent to find one. The bundled plan-with-corvi skill drives the loop end to end: read the plan, dispatch its todos in ordered file-disjoint waves, implement, then verify the diff against the plan it started from.
So Claude Code starts from a code-grounded plan instead of a blank prompt: the plugin can list and search your library, open a plan's Markdown and todos, dispatch the pending todos, and structurally check a git diff against the plan.
Reads are free; one tool can spend
Every tool but one only reads, and reading is local and $0 — it talks to the Corvi app on your Mac, not a server. The exception, create_plan, spends a small paid LLM call on your own key (~$0.01–0.05) to draft a fresh Quick Plan on demand.
create_plan ships available, and your agent still asks before every tool call — so nothing spends without your say-so. To make the plugin strictly read-only (headless or auto-approve setups), set the plugin's allow_create_plan config to "0", which passes CORVI_MCP_ALLOW_CREATE=0 to the server.Your plans stay local
The server the plugin runs is the Corvi app binary, reading the local plan store on your Mac. No plan text is uploaded; nothing is sent to Corvi to make the plugin work. The only thing that can leave is the paid model call behind create_plan, on your own key — the same opt-in egress posture as the app. See Egress & trust. Using a different MCP client? See Run as an MCP server.