Using genie in your harness

genie is an MCP server. Every harness that speaks MCP can call its 20 tools — ping plus 19 workflow tools — but how much guidance and GUI you get depends on two independent harness capabilities:

• Agent Skills — does the harness load a bundled SKILL.md? Claude, Cursor, Codex, GitHub Copilot, Continue, and Cline (via its own skill subcommand, confirmed 2026-07-14 — see cline.md) support the open Agent Skills format, with different install directories. The Skill teaches the conjure → plan → write_files → preview workflow so a plain-English request ("build me a button and show me") just works. Tool descriptions remain the fallback when a host or session does not load the Skill.
• ui:// rendering — can the harness render an MCP-Apps ui:// resource inline? If so, preview shows the card grid in-panel. If not, the genie server opens a browser tab only for a local stdio connection.

These are orthogonal, so harnesses fall into a grid:

Harness	Renders ui:// inline?	Loads Agent Skills?	What it relies on
Claude Code / Desktop / claude.ai	yes	yes	Skill + /genie:preview + tool descriptions
Cursor	yes	yes	Skill + inline grid + tool descriptions
VS Code Copilot (≥ Jan 2026)	capability-dependent	yes	Skill + inline grid when MCP Apps is negotiated + descriptions
ChatGPT remote connector	yes	no	inline grid + tool descriptions
Codex CLI	no	yes	Skill + server-opened browser tab + descriptions
GitHub Copilot (host-dependent)	capability-dependent	yes	Skill + inline grid when negotiated; local fallback browser + descriptions
Continue IDE	yes	yes	Skill + inline MCP App + descriptions
Continue CLI (cn)	no	yes	Skill + text fallback; headless cn -p; ${{ secrets.NAME }}
Cline	no	yes	Skill (cline skill add) + descriptions; local-stdio viewer/file://, remote HTTP degrades to text

Takeaways:

• On Claude Code, install the Skill + command (or the whole plugin) — see claude-code.md.
• On Cursor, Codex, and GitHub Copilot, install the same portable Skill using the path in each harness guide; descriptions remain a fallback.
• On ChatGPT's remote connector, tool descriptions carry the workflow and preview renders inline.
• On local stdio Codex / tools-only Copilot hosts, the genie server opens a browser tab itself when you call preview. HTTP defaults to remote locality and never auto-opens a browser. A genuinely same-machine HTTP client can opt into a manually opened local viewer URL with --preview-locality local (or GENIE_PREVIEW_LOCALITY=local). Remote HTTP hosts require the inline MCP App plus GENIE_PREVIEWS_BASE_URL. Disable local stdio auto-open with GENIE_PREVIEW_NO_OPEN=1.
• On Continue.dev, MCP tools only load in agent mode, not chat/autocomplete, and config secrets use ${{ secrets.NAME }} interpolation. Current Continue IDE builds include an MCP App renderer; the cn CLI remains text-only. Continue CLI loads Skills from .continue/skills, .claude/skills, and ~/.continue/skills. Current Continue infers transport when type is omitted; genie's snippets keep it explicit for readability — see continue.md.

Guidance delivery channels

1. Marketplace plugin — installs the Skill + namespaced command for an already registered genie MCP server in Claude Code. See claude-code.md.
2. Portable Skill copy — copy packages/plugin/skills/genie into the harness-specific Agent Skills directory documented below.
3. Tool descriptions — always present as fallback guidance when a Skill is unavailable or disabled.

Server prerequisites (all harnesses)

The current CLI validates its known secrets before starting any transport. GENIE_LLM_API_KEY is required at startup, including for read-only tool calls. OAUTH_HS256_KEY is optional and HTTP-only. GENIE_LLM_BASE_URL is required only for conjure and refine; the current conjure_screen implementation is an offline deterministic scaffold:

• GENIE_LLM_BASE_URL — the endpoint (must end in /v1).
• GENIE_LLM_API_KEY — its key (at least 16 characters).
• OAUTH_HS256_KEY — optional HTTP OAuth signing key (at least 32 characters when set).

Prefer environment variables or an owner-only mounted secrets file. Claude Desktop's manual local-server configuration is the exception: it has no separate secret prompt, so required values go in its owner-readable claude_desktop_config.json. Treat that file as sensitive and never commit it.

Per-harness registration snippets: claude-code.md · claude-desktop.md · cursor.md · codex.md · copilot.md · continue.md · cline.md