Skip to main content
Five minutes from zero to your coding agent calling lettuce — no CLI, no local setup.

​
1. Sign up

Create an account at uselettuce.dev/signup with email + password. The first person to sign in owns a new account; invite teammates later from the Team page.

​
2. Generate an API key

API keys (cwz_…) are what your coding agents present to lettuce — one per agent / machine / teammate is fine.
1

Open the API keys page

Go to uselettuce.dev/keys.
2

Click New key

Give it a recognisable name — e.g. claude-code-macbook or cursor-work-laptop.
3

Copy the key

The full cwz_… value is shown once, right after you create it. Copy it into your password manager now — you cannot retrieve it later.
Keys are stored hashed. A lost key cannot be recovered — revoke it from the same page and create a new one.

​
3. Add a repo

Open uselettuce.dev/repos, click Add repo, give it a name and a git URL. 
Name:   my-repo
Git URL: https://github.com/you/repo.git
Status moves pending → ready — usually within a minute. After that, the repo stays current automatically when you push.
Public repos only for now. Private-repo support is on the roadmap.

​
4. Connect your agent

Pick the snippet for the agent you use. All point at the same endpoint, https://diet.uselettuce.dev/mcp, with your cwz_… key in the Authorization header.

​
Claude Code

claude mcp add --transport http lettuce https://diet.uselettuce.dev/mcp \
  --header "Authorization: Bearer cwz_<your-key>"

​
Cursor

Save to .cursor/mcp.json in your project (or ~/.cursor/mcp.json globally): 
{
  "mcpServers": {
    "lettuce": {
      "url": "https://diet.uselettuce.dev/mcp",
      "headers": { "Authorization": "Bearer cwz_<your-key>" }
    }
  }
}
Full per-agent details (VS Code, Windsurf, Cline, mcp-remote) on the Connect your agent page.

​
5. Tell the agent to use it

Agents won’t reach for a new tool unprompted. Drop this into the repo’s CLAUDE.md (or your agent’s equivalent system prompt). This is the exact protocol Lettuce is benchmarked with — the discipline is what makes the graph save tokens instead of doubling them: 
# Codebase navigation — use the Lettuce MCP

This repo is indexed by Lettuce. For any "where is X / how does Y work /
who calls Z" question, use the Lettuce MCP instead of grepping.

1. Call `mcp__lettuce__understand(task='<the question>', repo='<repo>')`
   ONCE. It returns ranked candidate symbols with code, top caller/callee,
   docs hits, a `confidence` score, and a `next_step_hint`.

2. Branch on `next_step_hint.action`:
   - `answer_now`               → answer from the response. Don't grep,
                                  don't re-read "to verify" — the code
                                  returned is identical to what's on disk.
   - `answer_or_read_snippet`   → answer, unless you need one file fully —
                                  then call `mcp__lettuce__read_snippet` ONCE.
   - `read_snippet_then_answer` → call `read_snippet` ONCE on the top
                                  candidate, then answer.
   - `bail_to_grep`             → the graph missed; grep with the suggested
                                  `grep_terms`, then answer.

3. After at most 2 Lettuce calls (understand + at most one read_snippet),
   answer. Re-grepping or re-reading what `understand` already returned only
   adds cost, not correctness. If the graph genuinely missed it, say so
   rather than chasing with grep.
The exact location per agent (CLAUDE.md, .cursor/rules, .github/copilot-instructions.md, Windsurf rules, Cline custom instructions, …) is on the Agent instructions page. Ask: “Who calls process_payment, and what would break if I rename it?” — your agent should now call a lettuce tool to answer it.