AI-assisted teams

Let AI agents drive stacked pull requests

Give Claude Code, Cursor, and other MCP agents typed tools to create branches, submit stacks, and resolve conflicts — no fragile shell-outs.

The problem

AI coding agents are good at producing focused changes but bad at the mechanics of stacked git workflows over a shell. Parsing gt log output from stdout is brittle, and a mis-timed force-push can wreck a stack.

Typed tools, not shell parsing

stkd-mcp exposes gt_create, gt_submit, gt_restack, gt_sync, gt_land, and more as MCP tools with structured input and output schemas. The agent calls a tool and gets back typed data — no scraping terminal text, no guessing at exit codes.

The same guardrails as a human

The MCP server inherits the permissions of the user that runs it: it can only touch repos and remotes you already can. Pushing operations honor your provider's required reviews and branch protection, and a dry-run mode lets an agent explore without mutating anything.

Agent-shaped workflows

Because every CLI verb is a tool, an agent can compose the same workflows a human can: draft a feature as a stack, open the PRs, react to CI, and restack after edits — all inside one Claude Code or Cursor session.

The commands

  • cargo install stkd-mcp install the MCP server
  • gt_create MCP tool: start a tracked branch
  • gt_submit MCP tool: open the stack's PRs
  • gt_restack MCP tool: sync the stack

New to these? Follow the quickstart or read how stkd works.

Ready to stack your next change?

Install stkd in under a minute. Apache-2.0, single binary, every major platform.