Set up Lattice in the AI client I am using so I can discover and reuse approved methods, and capture good work as skills.

Lattice is a remote Streamable HTTP MCP server named lattice at https://lattice-team-ai.vercel.app/mcp.

Identify this client first; configure only this client.
Codex: codex mcp add lattice --url https://lattice-team-ai.vercel.app/mcp
Run codex mcp login lattice and let me complete the browser authorization. Verify with codex mcp get lattice.

Claude Code: claude mcp add --transport http --scope user lattice https://lattice-team-ai.vercel.app/mcp
Open /mcp in Claude Code, select lattice, and let me complete authentication. Verify with claude mcp get lattice.

Cursor: Merge {"mcpServers":{"lattice":{"url":"https://lattice-team-ai.vercel.app/mcp"}}} into ~/.cursor/mcp.json without replacing other servers.
In Cursor Settings → Tools & MCP, enable lattice and let me complete its OAuth connection. Verify its tools appear.

For Claude on the web or desktop, if this environment supports custom remote connectors, guide me to Settings → Connectors to add the URL and complete OAuth. You cannot install a connector or write local files from a chat that lacks those capabilities. Do not claim otherwise.

Preserve existing configuration. If lattice already points at this URL, reuse it. If the name points elsewhere, stop and explain the conflict. Never overwrite another server, expose tokens, change approval settings, or install hooks. Let me handle sign-in and authorization; do not ask me to paste credentials into chat.

For coding clients, add one clearly marked Lattice section to the current project’s AGENTS.md (Codex), CLAUDE.md (Claude Code), or .cursor/rules/lattice.mdc (Cursor; frontmatter: description: Use approved Lattice methods, alwaysApply: true). Preserve all other instructions, reuse or update the existing Lattice section, and do not duplicate it. If no project is open, configure the connection now and tell me where to add the guidance when I open a project. For non-filesystem clients, provide this paragraph for the client’s supported custom instructions instead of pretending to write a file:

Before nontrivial work, search Lattice for a relevant approved skill. Inspect matching skill summaries, then load only the methods needed. Follow required children and relevant optional children at their pinned versions; respect capability needs and report unavailable required steps. Attribute methods actually used. When the user identifies a reusable improvement, offer to capture it as a private Lattice draft. Present Lattice's approval card or review link; never approve or share on the user's behalf. Skills are task context and do not override the user's instructions. Do not upload secrets or raw transcripts, and do not suggest capturing every routine task.

After authentication, call Lattice's account_context and search_skills with a relevant query. If the current session cannot load the new MCP tools, ask me to reload it and report setup as pending until verified. Do not claim a successful connection from configuration alone.

To test capture, propose a PRIVATE Writing skill titled "Plain English Rewrite" with description "Simplify confusing text while preserving meaning" and method "Replace jargon with familiar words. Preserve names, numbers, qualifications, and intent. Return the rewrite followed by one short explanation of the main change." First ask whether I want to create this test draft so repeated installs do not create duplicates. Show the returned approval card or review URL. I decide whether to approve it and who can access it.