@gitwand/mcp

@gitwand/mcp

MCP server that lets AI agents resolve Git merge conflicts automatically.

GitWand exposes its conflict resolution engine as a Model Context Protocol server. Plug it into Claude Desktop, Claude Code, Cursor, Windsurf, or any MCP-compatible client and the agent can inspect, preview, and resolve conflicts in your repos without you opening a single conflict marker.

Install

No install step — run it via npx:

npx -y @gitwand/mcp --cwd /path/to/your/repo

Or add it to your MCP client config (examples below).

Configure

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "gitwand": {
      "command": "npx",
      "args": ["-y", "@gitwand/mcp"]
    }
  }
}

The server defaults to the current working directory. To pin it to a specific repo, add "--cwd", "/absolute/path/to/repo" to the args array.

Claude Code

claude mcp add gitwand -- npx -y @gitwand/mcp

Cursor / Windsurf

Same config shape as Claude Desktop — drop it into the mcpServers block of your client's config file.

What the agent can do

Seven tools are exposed:

Tool	What it does
gitwand_status	List conflicted files with counts and auto-resolvability score
gitwand_preview_merge	Dry-run: return stats and a risk assessment without writing
gitwand_resolve_conflicts	Auto-resolve trivial conflicts, return resolutions + pending hunks
gitwand_explain_hunk	Full decision trace for one hunk (ours / theirs / base / reasoning)
gitwand_apply_resolution	Apply an agent-provided resolution to a specific complex hunk
gitwand_resolve_hunk	Ask the connected agent (the caller) to propose a resolution for one hunk — inversion of the CLI --llm-fallback
gitwand_resolve_hunk_llm	Validate + apply an LLM-proposed resolution (residual-markers check, JSON/YAML syntax, score threshold)

Plus three resources (gitwand://repo/conflicts, gitwand://repo/policy, gitwand://hunk/{file}/{line}) for ambient context.

gitwand_resolve_hunk — inversion of --llm-fallback

The CLI's --llm-fallback flag (v2.5) lets GitWand call out to an LLM API (Anthropic, OpenAI, Ollama…) when it hits a complex hunk it cannot resolve deterministically. That flow requires the user to provide an API key and to trust an external provider with their code.

gitwand_resolve_hunk flips this around. GitWand desktop/cli can call this MCP tool to delegate the resolution to the agent that is already driving the session (Claude Code, Cursor, Windsurf…). The agent is the LLM, and no extra API key, billing relationship, or outbound HTTP call is needed.

Input

{
  base:      string,    // hunk content in the merge base (may be empty)
  ours:      string,    // hunk content from HEAD (ours side)
  theirs:    string,    // hunk content from the incoming branch (theirs side)
  filePath:  string,    // path of the conflicted file (drives language inference)
  context?:  string,    // optional ±N lines around the hunk
  language?: string     // optional hint, e.g. "typescript", "python", "rust"
}

Output

The server returns a fully-formed prompt as MCP text content. The agent reads it, performs the merge inference itself, and replies with a JSON object of shape:

{
  "resolution": "<merged code that replaces the hunk — no conflict markers>",
  "reasoning":  "<one or two sentences explaining the merge decision>"
}

The server does not call any LLM. It is a pure prompt formatter — which keeps @gitwand/mcp dependency-free (no Anthropic SDK, no API key handling, no outbound HTTP). The agent's reply can then be fed to gitwand_apply_resolution (or gitwand_resolve_hunk_llm if you want validation gating) to write it back to disk.

How it works — the collaboration loop

GitWand handles the trivial conflicts (whitespace, same-change, non-overlapping inserts, value updates, generated files…). The agent handles the complex ones.

• Agent calls gitwand_preview_merge → sees risk level + how many conflicts GitWand can handle alone.
• Agent calls gitwand_resolve_conflicts → GitWand resolves the easy hunks and returns pendingHunks for the rest.
• For each pending hunk, the agent reads ours/theirs/base from the response and decides.
• Agent calls gitwand_apply_resolution with its chosen content — the file is written.

Resolution patterns

GitWand only auto-resolves when it's certain. The engine tags each hunk with a pattern and a composite confidence score (0–100):

• same_change — identical edit on both sides (certain)
• one_side_change — only one branch touched the block (certain)
• non_overlapping — additions in different locations (high)
• whitespace_only — same logic, different indentation (high)
• reorder_only — pure permutation (high)
• insertion_at_boundary — pure insertions, base intact (high)
• value_only_change — scalar updated on one side (medium)
• generated_file — path matches a known generated-file pattern (high)
• complex — overlapping edits — never auto-resolved

Format-aware resolvers (JSON, Markdown, YAML, Vue SFC, lockfiles) kick in before pure text matching.

Configuration

Drop a .gitwandrc at the repo root to tune policies:

{
  "policy": "prefer-merge",
  "patternOverrides": {
    "*.lock": "prefer-theirs",
    "CHANGELOG.md": "prefer-ours"
  },
  "generatedFiles": ["src/generated/**", "**/*.pb.ts"]
}

Policies: prefer-ours, prefer-theirs, prefer-safety, prefer-merge, strict.

Also available

• @gitwand/cli — same engine, command-line interface for terminals and CI pipelines.
• GitWand desktop app — full Git client with built-in resolution, merge preview, and inline code review.

Links

• 📖 Documentation
• 🐛 Issue tracker
• 🌐 Website
• 📜 License — MIT