Skip to content

AI-Assisted Development with MCP#

Connect your AI coding assistant to the live Jac compiler in five minutes. The built-in jac mcp server exposes ~19 tools (validate, lint, run, transpile, explain errors, search docs), the bundled reference guides, and every example project over the Model Context Protocol -- so your assistant stops guessing at Jac syntax and starts checking its work against the real compiler.

What you'll do:

  1. Start the MCP server and inspect what it offers
  2. Connect Claude Code (or any MCP client)
  3. Use it: a write → validate → run loop where the assistant verifies its own Jac

Time: ~10 minutes. Nothing to install -- the server ships inside the jac binary with zero third-party dependencies.

1. See what the server offers#

jac mcp --inspect

This prints the full inventory and exits: 51 resources (the bundled guides like jac://guide/jac-walker-patterns, the grammar spec, and complete example apps), 19 tools, and 9 prompts (scaffolds like write_walker and fix_type_error). The tools your assistant will lean on most:

Tool What it does
validate_jac Full type-checked validation with structured errors
check_syntax Fast parse-only check
run_jac Execute code and return stdout/stderr
lint_jac / format_jac Style checks with auto-fix
explain_error Root cause + fix example for a compiler error
jac_to_py / jac_to_js Show the generated Python / JavaScript
search_docs Ranked snippets from the guides and reference
graph_visualize Run code and return the resulting graph as DOT/JSON

2. Connect your client#

For Claude Code, one command:

claude mcp add jac -- jac mcp

For Claude Desktop, Cursor, Windsurf, or VS Code, add the standard stdio-server block to the client's MCP config (exact file locations and copy-paste snippets for each client are in the MCP Server reference):

{
  "mcpServers": {
    "jac": {
      "command": "jac",
      "args": ["mcp"]
    }
  }
}

Restart the client and the Jac tools appear in its tool list. If your model is small or its context is tight, register the server as jac mcp --mode lite to expose a reduced tool set.

3. Use it -- the verify loop#

Open a conversation in your connected assistant and try:

"Write a Jac walker that counts all nodes reachable from root, validate it with the jac tools, and run it on a small example graph."

Watch the tool calls: the assistant drafts code, calls validate_jac, gets structured type errors back (not a guess -- the actual compiler), fixes them, then run_jac executes the result and returns real output. Two habits make this loop reliable:

  • Ask for validation explicitly at first ("validate with the jac tools before showing me code"); most assistants then adopt the habit for the rest of the session.
  • Point it at the guides for idioms: "check jac://guide/jac-walker-patterns first" -- or let it discover them with search_docs.

When the assistant hits a compiler error it can't parse, explain_error returns the error's category, root cause, and a fix example -- the same content jac check links to in your terminal.

Beyond MCP#

  • Auto-loaded skills instead of a server: jac guide --export ~/.claude/skills writes the same guides as Agent Skills that Claude Code and Cursor load automatically -- see Agent Skills and MCP for when to use which.
  • Jac's own agent: jac ai is a built-in coding agent that already knows the guides and runs against local models with no API key -- see the jac ai reference.
  • Full tool catalog, transports, troubleshooting: the MCP Server reference.