Agent Skills and MCP#

AI coding assistants are good at Jac's ideas but often wrong about its syntax -- the language has evolved, and models routinely confuse Jac with Python or JSX. The jac CLI ships the corrective reference built in, so there is nothing to install.

jac guide -- curated reference guides bundled with the compiler. They are the authoritative spec for writing correct, idiomatic Jac, and any agent that can run a shell command can read them.
The jac mcp server -- a Model Context Protocol server that gives your assistant live compiler tools: validate, format, lint, run, transpile, and search the docs. It also serves the same guides as MCP resources.
jac ai -- Jac's own built-in coding agent, if you'd rather not bring an external assistant at all. It knows the guides already and works with fully local models, so it runs without an API key.

These are complementary: the guides tell a model how Jac works; MCP lets it verify what it wrote against the real compiler; jac ai packages both into a ready-made agent.

`jac guide` -- the built-in reference#

The guides ship inside the jac CLI -- one per topic (jac-core-cheatsheet, jac-types, jac-walker-patterns, jac-by-llm, the jac-sv-* server guides, the jac-cl-* client guides, and more). They are always version-matched to the compiler you have installed.

jac guide                      # list every available guide
jac guide jac-types            # print a specific guide
jac guide --search walker      # find guides by keyword
jac guide --json               # machine-readable list (for tools and agents)

Because the guides are part of the CLI, an AI agent working in your project can self-serve them with no setup -- it just runs jac guide. Two things reinforce this:

jac create seeds an AGENTS.md in every new project, telling agents to consult jac guide.
jac check diagnostics link to guides. When the type checker flags an error it points at the relevant guide -- e.g. a type error prints → run 'jac guide jac-types' for guidance -- so the model is pulled to the fix at the moment it is wrong.

Export as Agent Skills (Claude Code, Cursor)#

Claude Code, Cursor, and the Claude Agent SDK can auto-load Agent Skills -- reading a skill's frontmatter and pulling in the full guide when a task matches. To get that automatic behaviour, export the guides as a skills directory:

Personal (all projects)Project (this repo only)

jac guide --export ~/.claude/skills

jac guide --export .claude/skills

Commit .claude/skills/ to version control so everyone on the project gets the same Jac guidance.

jac guide --export writes each guide as <dir>/<name>/SKILL.md -- the dir-per-skill layout Claude Code and Cursor discover. Re-run it after upgrading Jac to refresh the guides. The same command works for Cursor (jac guide --export ~/.cursor/skills).

Export is optional

You only need to export if you want automatic skill loading. Any agent that can run jac guide already has the full reference on demand.

MCP server (any MCP client)#

The built-in jac mcp server runs a Model Context Protocol server that exposes the Jac compiler -- grammar, documentation, examples, the bundled guides, and tools to validate, format, lint, run, and transpile Jac -- to any MCP-capable assistant.

Start it with:

jac mcp

Then register it with your client. For Claude Code:

claude mcp add jac -- jac mcp

Other clients (Claude Desktop, Cursor, Windsurf, VS Code) use a JSON configuration block. The MCP Server reference has copy-paste configuration for every supported client, the full tool and resource catalog, transport options, and troubleshooting.

Tip

The MCP server is built into the jac binary -- there is nothing to install. Just run jac mcp. (It has no third-party dependencies; the protocol is implemented on the Python standard library.)

`jac ai` -- the built-in coding agent#

If you want an agent now, without wiring up an external assistant, the CLI ships one:

jac ai                                   # interactive session (project's configured model)
jac ai "add a walker that lists todos"   # one-shot request
jac ai -m local:gemma-4-e4b              # fully local -- no API key
jac ai --ui                              # web UI with a live phase-graph visualizer

It uses your [byllm.model] configuration (falling back to the bundled local model), reads the same built-in guides, and can edit files and run code in your project -- pass --safe to approve every write and command. See the jac ai reference.

Structured code access for agents#

Two more commands exist mainly to make your project legible to agents (yours or an external one):

jac code -- compiler-backed structural queries (symbol, uses, map, walkers, slice, diag) that return JSON. An agent can ask "which walkers touch Todo nodes?" instead of grepping. See the jac code reference.
jac browse -- headless-browser automation over CDP (navigate, click, snapshot, screenshot), so an agent can drive and visually verify the web app it just built. See the jac browse reference.

Which to use#

	`jac guide`	`jac guide --export`	MCP (`jac mcp`)
Provides	Reference knowledge, on demand	Auto-loading Agent Skills	Reference + live compiler tools
Discovery	Agent runs the CLI	Assistant loads by frontmatter	Client lists MCP resources/tools
Setup	None -- built in	One `--export` command	Run a server, register it
Best for	Any agent that runs a shell	Claude Code, Cursor, Agent SDK	Any MCP client; verifying code

For the strongest setup, export the guides so your assistant writes idiomatic Jac, and connect jac mcp so it can validate and run what it writes against the real compiler before handing the code back to you.