MCP for coding agents

sprout mcp [root] speaks the Model Context Protocol over stdio. Agents get a map of the project in one call, instead of spending their context on ls -R, grep and reading files one by one.

Set it up

Sprout can print the setup for your client with its own full path filled in:

sprout mcp --print-config claude-code     # or cursor, vscode, claude-desktop

Or copy it from here.

Claude Code: install the plugin. It adds the server, and a skill that tells Claude when Sprout answers better than grep:

/plugin install sprout --marketplace Sprout-DevLabs/sprout

It runs the sprout on your PATH. Or add just the server, in your project:

claude mcp add sprout -- sprout mcp

Use one or the other: with both, Claude sees every tool twice (claude mcp remove sprout removes the server).

Cursor: .cursor/mcp.json in your project, or ~/.cursor/mcp.json for every project. Cursor fills in ${workspaceFolder}:

{ "mcpServers": { "sprout": { "command": "sprout", "args": ["mcp", "${workspaceFolder}"] } } }

VS Code: .vscode/mcp.json in your project. The top-level key is servers, not mcpServers:

{ "servers": { "sprout": { "type": "stdio", "command": "sprout", "args": ["mcp", "${workspaceFolder}"] } } }

Claude Desktop: Settings → Developer → Edit Config opens ~/Library/Application Support/Claude/claude_desktop_config.json on macOS, or %APPDATA%\Claude\claude_desktop_config.json on Windows. Desktop apps don’t know your project, and often can’t find Homebrew’s binaries, so give both in full. which sprout prints the first:

{ "mcpServers": { "sprout": { "command": "/opt/homebrew/bin/sprout", "args": ["mcp", "/path/to/your/project"] } } }

Restart the client after editing its config.

Which folder it serves

1. The folder you pass: sprout mcp /path/to/project.
2. Otherwise, the folder your client says is open, if it supports MCP roots (VS Code and Claude Code do). When the client says it changed, the server follows.
3. Otherwise, the folder it was started in.

In / or your home folder, whether the server started there or the client is open there (Claude Code launched from ~, say), the tools refuse and say how to fix it, instead of mapping hundreds of thousands of files, private ones included. Start the agent inside the project, or pass the project path. Pass the home folder explicitly if you really mean it.

Any single call can also name its project with root, an absolute path. The agent can then work on a repository other than the open folder, even while that folder is refused.

Tools

Tool	Arguments	Returns
project_map	path?, budget?	The --ai map with key files, at a 1000-token budget unless the agent asks for more. Its description tells agents to call it first
reading_order	path?	The --entry list
tree	path?, depth? (default 3), all?, git?, churn?, since?	A tree of a folder
diff_tree	rev, path?, depth?	The --diff view
dependents	file, depth?, noTests?	What depends on a file, and why (dependents)
deps	file, depth?	What a file depends on, and why
impact	files? or staged?, rev?, commit?; all?, symbol?	What a change could break, and the tests to run (impact). Uncommitted changes by default; symbol with one Go file narrows it to one function, method, type or variable
context	file, budget?	What to know before editing a file (context)

Every tool also takes root?, the project’s absolute path, for that call only.

The tool descriptions tell agents when to reach for each: project_map first in a new repository, context or dependents before editing a file, impact after editing.

Safety

An agent’s arguments are untrusted input, so:

• path, file and files must stay inside the root once .. and symlinks are resolved. Absolute paths and URLs are refused, so the server never clones anything.
• root must be an absolute path to a project folder (one with .git or a manifest like go.mod or package.json). It can’t be a hidden folder, / or your home folder, and if you gave the server a folder, it must be inside that folder.
• rev and commit can’t become git options.
• Output uses paths relative to the root. The one exception: when a file isn’t found, the error names the folder being served, because a wrong folder is the usual cause.
• The server is read-only.

If it doesn’t work

• No sprout tools show up: restart the client. Then run sprout mcp in a terminal: it should print MCP server on stdio, root … and wait. If the client’s MCP log says the command wasn’t found, it can’t see your PATH: use the full path from which sprout.
• The map is of the wrong folder: start the agent inside the project, or pass the project path as in the examples above. When a file isn’t found, the error says which folder the server is on.
• The agent doesn’t use the tools: the server tells agents when each tool beats searching by hand, but asking helps: “use sprout’s impact tool on what you changed”.
• Stale answers: there aren’t any. Every call reads the disk again, so files your agent just wrote are in the next answer.