How do I install stkd-mcp with Claude Code?

Run cargo install stkd-mcp and add { "mcpServers": { "stkd": { "command": "stkd-mcp" } } } to your Claude Code config (claude_desktop_config.json or .mcp.json). Restart Claude Code and the stkd tools appear.

Which tools does stkd-mcp expose?

gt_create, gt_submit, gt_restack, gt_sync, gt_land, gt_log, gt_modify, gt_continue, gt_undo. The full surface mirrors the CLI, so agents can compose any workflow a human can.

Is the MCP server safe to give to autonomous agents?

stkd-mcp inherits the permissions of the user that runs it — it can touch only the repos and remotes you can already touch. Operations that push code (gt_submit, gt_land) honor your provider's required reviews and branch protection. Run in dry-run mode (--check) for non-mutating exploration.

MCP integration

Stacked diffs, driven by AI agents.

stkd-mcp is a Model Context Protocol server that exposes every gt command as a typed tool. Drop it into Claude Code, Cursor, or any MCP-capable agent and the model can structure work as a stack — natively.

Quickstart

1. Install stkd-mcp

cargo install stkd-mcp

2. Wire it into Claude Code

Add this to ~/.claude.json (or claude_desktop_config.json):

{
  "mcpServers": {
    "stkd": {
      "command": "stkd-mcp"
    }
  }
}

3. Restart Claude Code

You should see stkd appear in the tools menu with the sub-tools below.

The tool surface

Each tool is a typed function call. JSON Schema input and output. Agents compose them freely.

gt_create

Create a branch stacked on top of the current branch.

gt_log

Return the current stack as JSON: branches, parents, current pointer, PR numbers.

gt_modify

Amend the current branch's tip commit with new changes.

gt_restack

Rebase every dependent branch on top of its parent's latest tip.

gt_submit

Push branches and create or update PRs / MRs. Supports --stack to submit the whole stack.

gt_sync

Fetch origin, detect merged branches, clean up, and restack.

gt_land

Merge the bottommost open PR; with --stack, merges every PR in order.

gt_continue

Resume an in-progress restack after a conflict has been resolved.

gt_undo

Roll back the last stkd mutation. Conservative — it remembers the last clean state.

Safe by default

stkd-mcp can do anything you can do — including pushing code. We assume that's intentional, but we also ship four guardrails for restricting agent power:

Allow-list tools

"args": ["--tools=gt_log,gt_create"]

Only the listed tools are exposed. Useful for "let the agent plan but not act."

Dry-run

"args": ["--dry-run"]

Every mutating tool becomes a no-op that reports what would have happened.

Branch pattern lock

"args": ["--branch-pattern=claude/*"]

The agent can only create branches matching the pattern. Your protected branches stay protected.

Repo allow-list

"args": ["--repo=~/code/safe/*"]

Only repos matching the glob are exposed to the agent.

What agents actually do with it

Split a big diff into a stack. Agent reads the working tree, identifies cohesive groups, lays out the stack, and submits.
Sweep restacks. After main advances, the agent iterates over your open stacks, runs gt_sync, and reports which ones now have conflicts.
Conflict triage. Agent reads the conflict markers, attempts trivial resolutions (imports, renames, formatting), and hands non-trivial ones back with a clear summary.
CI babysitter. After gt_submit, the agent polls PR statuses via gh / glab and surfaces failures the moment they happen.

Full walkthrough in Using stkd with Claude Code.

MCP FAQ

Agent-integration questions.

stkd-mcp exposes every stkd command as an MCP tool, so AI agents like Claude Code and Cursor can manage stacked pull requests without shelling out to the CLI. The agent sees structured input/output schemas and can chain operations safely.

Wire it up in under five minutes.

cargo install stkd-mcp, add 5 lines of JSON, restart Claude. That's the whole setup.

Install stkd-mcp → See an example session