Your AI agents, with a real Git engine on tap.
GitWand's MCP server exposes its deterministic conflict engine to AI agents. GitWand resolves the trivial 95% instantly and deterministically; your agent takes over the complex cases with full context — ours/theirs/base content, classification trace and confidence scores. AI is always opt-in, always auditable.
How an agent uses GitWand
Three steps: the engine handles the deterministic work, the agent handles the judgment calls.
Preview
The agent calls gitwand_preview_merge to assess the number of conflicts, their complexity, and the percentage GitWand can resolve on its own.
Auto-resolve
GitWand instantly resolves trivial patterns (whitespace, one-side-change, same-change…) and returns ambiguous hunks with their classification trace.
AI resolution
For each complex conflict, the agent has full context: ours/theirs/base content, classification trace and confidence scores — then proposes a resolution you can accept or reject.
One command to wire it up
With Claude Code: claude mcp add gitwand -- npx -y @gitwand/mcp. For Claude Desktop, Cursor or Windsurf, drop the config block into your client. The server is also on the official MCP Registry, so registry-aware clients discover it automatically.
{
"mcpServers": {
"gitwand": {
"command": "npx",
"args": [
"@gitwand/mcp",
"--cwd",
"/path/to/repo"
]
}
}
}LLM fallback, when you get stuck
For the hunks patterns can't resolve, opt in to LLM resolution via Claude, OpenAI-compatible endpoints, Ollama, or MCP — your model, your key. Every suggestion runs through the same parse-tree validation pipeline as the deterministic patterns, with a decision trace and a reject button. No built-in model, no data leaving your machine unless you choose it.
Opt-in only
Deterministic patterns run first and free. The LLM is called only for what's left.
Validated
LLM output is parsed and validated like any other resolution — not trusted blindly.
Your provider
Claude · OpenAI-compatible · Ollama · MCP. Your key, your endpoint, your data.
Auditable
Full decision trace and a one-click reject on every proposed resolution.