Agent integrations

Slide Agent publishes one composition language. Every integration below is a different way of delivering it to a model — the model still does the designing in all of them.

Run slide-agent doctor to see which are live on your machine. It reports registered (Slide Agent wrote the skill where it expects the host to look) separately from verified (a host configuration actually references it), because only the second is evidence.

Support levels

Target	Level	Mechanism
Codex / Agent Skills	verified	~/.agents/skills/slide-agent
Claude Code	verified	~/.claude/skills/slide-agent
MCP clients (Cursor, Zed, Windsurf, …)	verified	slide-agent-mcp over stdio
CLI / any tool-capable agent	verified	slide-agent catalog + slide-agent build
GitHub Copilot CLI	best-effort	~/.copilot/skills/slide-agent
Gemini / Google Antigravity	verified	Global plugin at ~/.gemini/config/plugins/slide-agent-plugin, with the skill under skills/slide-agent
VS Code extension	verified	Uses the VS Code Language Model API

We label these honestly rather than claiming universal support. If a level here is wrong in either direction, please open an issue.

The gemini installer target creates the system-wide user plugin layout:

~/.gemini/config/plugins/slide-agent-plugin/
├── plugin.json
└── skills/
    └── slide-agent/
        └── SKILL.md

MCP

{
  "mcpServers": {
    "slide-agent": { "command": "slide-agent-mcp" }
  }
}

Seven tools — slides_catalog, slides_build, slides_edit, slides_view, slides_finalize, slides_inspect, slides_generate — with the composition grammar, the catalog, the intent schema, every recipe, and worked examples as resources. A client that has never heard of Slide Agent can learn the language at runtime by reading slide-agent://grammar, which is about 1,300 tokens.

Each tool has a response-token budget, so a build verdict cannot flood a context window: catalog ≤ 3,000, build ≤ 800, edit ≤ 500.

Adding --compat-v1 (or SLIDE_AGENT_COMPAT_V1=1) registers the 0.x surface — 9 tools, 21 resources, 2 prompts — alongside the seven. That is available for one minor release and goes away in 2.1.

Full MCP reference → — per-client configuration, the call sequence that produces good decks, every tool's arguments, the resource URI scheme, and error codes.

Any CLI-capable agent

No integration required:

slide-agent catalog --grammar                              # the language, as a system prompt
slide-agent catalog --include schema                       # JSON Schema for structured output
slide-agent build --intent deck.intent.json --deck out/    # build it
slide-agent view --deck out/ --what sheet                  # look at it

This is the fallback that always works, including for self-hosted models.

Self-hosted models

Fetch the schema, request structured output against it, and build what comes back. Two things make this practical for a smaller model than you might expect: the grammar is short, and every finding carries the JSON pointer it applies to — so a failed build hands back an addressed list of what to change rather than prose to interpret.

If the model cannot hold the whole grammar, start it on recipes (slide-agent catalog --include recipes) and let it expand the ones it wants to rework. Every verdict reports how much of the deck came from recipes, so the tradeoff stays visible.

The one mode that calls a model itself

slides_generate / slide-agent generate is the exception to everything above: it runs a director and a critic inside Slide Agent, for callers who have no model of their own. It needs ANTHROPIC_API_KEY and the optional @anthropic-ai/sdk peer dependency. Nothing else in Slide Agent makes a network call to a model.