Lusha MCP Plugin

Find and enrich B2B contacts and companies with verified emails, direct dials, mobile numbers, and real-time buying signals from Lusha — straight from inside your AI assistant.

Supports Codex (plugins), Claude Code (Claude Code CLI / Cowork), Cursor (plugins), VS Code Copilot (GitHub Copilot Chat with MCP), and Gemini CLI (extensions).

Skills

Find and enrich

Prioritise and follow up

Write outreach — a two-stage pair, run in order

How it works

The find-and-enrich skills each chain multiple Lusha API calls into a complete workflow, and surface verified phone numbers prominently — direct lines and mobile numbers are first-class outputs, not an afterthought.

The prioritise skills start from what you already have: your ICP profile, your own list, or your website traffic. Each one starts with the free calls (your ICP profiles, your CRM connection, your visit data) and branches on what it finds. Anything not set up yet gets a setup offer or a fallback, not an error.

Every skill that spends credits:

• states the cost before the call
• previews 10 before pulling a full page
• reveals only the people you pick

Some reveals finish in the background when third-party sources answer later. The skills wait for those reveals rather than charging twice. Each find, prioritise and enrich skill ends with the same two options, save to a Workspace table or export to your CRM (both through save-or-export), plus one related skill.

Every Lusha call a skill makes passes the conversation_id from the previous response and tags reason_for_invocation with the skill's name.

The outreach pair sits on top of them. outreach-research is pure conversation, producing a brief you save and reuse. outreach-sequence consumes that brief and never calls Lusha directly: it delegates signal discovery and harvest to signal-prospect, and contact enrichment to enrich-contact, so credit handling stays in one place. Every call it triggers is tagged with a reason_for_invocation starting outreach-sequence: , which is how skill-driven usage is attributed (each other skill tags its calls with its own name the same way).

All clients load the same skills/*/SKILL.md files and the same Lusha MCP server — only the per-client manifest and store endpoint differ:

Skills reference Lusha tools by their bare logical name (e.g. contacts_search), so a single skill source works identically across all clients. Gemini CLI auto-discovers the bundled skills/ directory as extension skills.

Prerequisites

• A Lusha account with API access

Install

Codex

The Codex plugin lives at the repo root — .codex-plugin/plugin.json (manifest) and mcp.json (MCP server), with skills: "./skills/" pointing at the shared root skills/. Codex discovers it through the repo marketplace catalog at .agents/plugins/marketplace.json, which uses a url source pinned to a branch/tag. That catalog is read only by Codex/OpenAI tooling — Claude, Copilot, and Gemini keep using their own provider-specific manifests.

A url source is used instead of a local path because Codex rejects a local plugin path that resolves to the repo root (codex#17066) and silently drops symlinks during install (codex#18863). Cloning the repo over url keeps skills/ as real files at the plugin root, so no copy or symlink is needed.

Add the marketplace and install:

codex plugin marketplace add lusha-oss/lusha-mcp-plugin
codex
/plugins

Select Lusha Plugins, install the Lusha plugin, then start a new Codex thread so the skills and MCP tools are loaded. The url source installs from the ref pinned in .agents/plugins/marketplace.json, so changes take effect once they land on that ref.

Claude Code (CLI / Cowork)

/plugin marketplace add lusha-oss/lusha-mcp-plugin
/plugin install lusha

Cursor

Cursor reads the plugin manifest at .cursor-plugin/plugin.json and discovers the bundled skills/ automatically. Add the repo as a plugin marketplace, then install the Lusha plugin from .cursor-plugin/marketplace.json (catalog lusha-plugins, plugin lusha). All skills activate from natural-language requests once the MCP server connects.

VS Code Copilot

Requires a VS Code version with agent-plugin support and the GitHub Copilot extension. The plugin bundles the MCP server and all skills together via .github/plugin/plugin.json.

1. Open the Command Palette (Cmd+Shift+P / Ctrl+Shift+P).
2. Run Chat: Install Plugin From Source.
3. Paste the repository name: lusha-oss/lusha-mcp-plugin.

The Lusha MCP server and all skills load automatically. Invoke a skill from Copilot Chat with its name, e.g. /enrich-contact, /prospect, /work-my-accounts, /score-my-list, or /outreach-sequence.

Gemini CLI

The repo ships a gemini-extension.json manifest at its root, so Gemini CLI wires up the Lusha MCP server and discovers the bundled skills automatically.

gemini extensions install https://github.com/lusha-oss/lusha-mcp-plugin

All skills are registered as extension skills — Gemini activates the matching one on demand (e.g. when you ask it to find a contact's phone number, build a prospect list, or draft outreach). Run gemini skills list to confirm they loaded.

Authentication

The Lusha MCP server uses OAuth. The first time you invoke a Lusha skill or tool, you'll be prompted to sign in with your Lusha account. Subsequent calls reuse the authenticated session.

Skill chaining

Skills are designed to feed into each other:

• prospect → signal-prospect: build a list, then filter it to companies showing buying signals
• lookalike-prospect → signal-prospect: find lookalikes, then prioritize by signal
• enrich-contact → lookalike-prospect: enrich a single contact, then find similar people
• prospect / lookalike-prospect → score-my-list: rank the list you just built by ICP fit and buying signals
• work-my-accounts ↔ set-up-my-icp: recommendations run on a calculated profile until you set up your own; a new profile previews its top accounts
• score-my-list → work-my-accounts: find the decision makers at your A tier
• any find or prioritise skill → save-or-export: save the result to a Workspace table or export it to your CRM
• outreach-research → outreach-sequence: capture positioning once, then draft against it
• prospect → outreach-sequence: hand a lead list straight into drafting (optional — any CSV with full_name, company, and title works)

outreach-sequence requires a brief from outreach-research; without one it refuses to draft rather than inventing positioning. An audience is required too, but it can come from anywhere: a pasted CSV, an attached file, a single named contact, or a prospect handoff.

Interactive and automated runs

outreach-sequence runs in two modes, set out in its Execution Mode section:

• Interactive (default) — a human is in the conversation, so the skill asks about signal preferences, any audience-wide signal, and whether you want a sample before it drafts the batch.
• Automated — the caller declares a non-interactive run ("non-interactive", "headless", "automated run", or "no user is available to answer"), as an automation runner or scheduled job would. Every question resolves from a supplied value or a documented default and execution continues, because an emitted question would become the final output and the run would produce nothing.

Automated mode never relaxes the data rules: the brief and audience stay hard-required, signal and enrichment work still delegates to the other skills, and nothing may be fabricated to fill a gap.

What the drafts follow

• Email 1 is five short parts (why now, pain, what the product does, one proof point, one question), 50–90 words, with a 3–5-word subject. It asks about interest, never for a meeting, unless you ask for one.
• A signal is the reason for writing only when it is relevant to the persona's pain, dated within 90 days, and stated as the source states it: the month, never "last week"; no percentages or estimated figures. Website visits and intent topics set the timing but never appear in a message. When no signal qualifies, Email 1 leads with the persona's pain and names one real detail about the contact or their company.
• Email 2 brings one new reason (a second signal, a proof point, or a question about the pain); Email 3 closes the thread by naming Email 1's topic. Both reuse Email 1's subject so your sending tool threads them.
• LinkedIn copy is the message for once they accept; send the connection request blank unless you ask for notes (at most 200 characters).
• No placeholders. A part that needs a fact the brief lacks is left out, and one line under the table says what to add to the brief. A personal touch comes only from a Lusha enrich result that shows a prior employer: the prior title and employer, nothing more.

Output columns (0.1.3)

linkedin_message (the message for once they accept) and linkedin_connection_note replace linkedin_message_1–linkedin_message_3, and every drafted row gains hook_source and personal_touch_source. Update any Braze, Outreach or Salesloft template that used the old LinkedIn columns.

Brief format

outreach-research writes brief v2 (<!-- lusha-outreach-brief v2 -->): it adds proof points (each marked OK to name in email), titles to match per persona, an optional email style, and an (unconfirmed) marker for values from a web search that you have not confirmed. Both skills still read a v1 brief.

Contributing

Root skills/ is the single source of truth — every client (Codex, Claude, Copilot, Gemini) reads these same files, so edit skills only under skills/.

When releasing, update the ref in .agents/plugins/marketplace.json to the branch or tag Codex users should install from (e.g. master for production, or a feature branch while testing).