iGPT Skills

Email intelligence for every role in the company. A Claude Code plugin marketplace that turns the email already in your inbox into structured, queryable insight — by role, by skill, in plain language.

13 role-specific plugins · 76 focused skills · 13 agents · all powered by the iGPT MCP at https://mcp.igpt.ai/.

What this is

iGPT Skills is a marketplace of Claude Code plugins that lets people ask plain-language questions about their email and get back grounded, structured answers — by role.

Each plugin targets a specific role (sales, finance, HR, ...) and ships:

• A tightly-scoped set of skills that each answer one focused question.
• An agent that handles broad, cross-skill questions and synthesizes a prioritized response.
• An MCP configuration that points at the iGPT backend at https://mcp.igpt.ai/.

A salesperson asking "who's waiting on a reply from me?", a finance lead asking "build me an expense report for May in USD", a researcher asking "what grant deadlines are coming up?", or a CSM asking "which customers should I be worried about?" — the right skill picks itself up and returns a clean, ready-to-act answer grounded in their actual email threads.

Who it's for

End users. Salespeople, customer success managers, finance leads, HR business partners, recruiters, executives, marketers, real estate agents, consultants, procurement leads, IT teams, researchers — anyone whose work runs through email.

Builders and integrators. Developers extending the plugins, customizing prompts and schemas, or integrating the iGPT API directly. Every skill's prompt and structured output schema is documented at the file level (plugins/<plugin>/skills/<skill>/SKILL.md) and reusable.

Quick start

Install one of the plugins and connect your email through iGPT. That's the whole setup.

Cowork (Claude desktop app)

1. Open the plugins section in Cowork. Search the marketplace for iGPT, or add from GitHub: paste igptai/skills (or the full URL https://github.com/igptai/skills).
2. Pick the plugin matching your role (e.g. igpt-sales).
3. Install. The first time you ask the plugin a question, you'll be prompted to sign in to iGPT (free account; OAuth in a browser; read-only email permissions).
4. Ask in plain language: "who's waiting on a reply?", "what grant deadlines are coming up?", "any compliance flags I should know about?"

Claude Desktop

Settings → Connectors → Add MCP Server → enter https://mcp.igpt.ai/ and complete OAuth.

Claude Code

/plugin marketplace add igptai/skills
/plugin install igpt-sales@igpt-skills   # or any plugin name

The plugin ships its own .mcp.json so the MCP server is registered for you. OAuth happens on first use.

Other MCP clients

Any Streamable-HTTP MCP client works. Point it at https://mcp.igpt.ai/ and use the prompts and JSON schemas from each SKILL.md directly.

At a glance

Plugins

Every plugin has a dedicated README.md (linked above) with chat-style examples, a "How it works" pipeline, and hand-held install steps tuned to the role. The skill tables below give a one-line summary of what each skill solves.

igpt-sales — Sales intelligence

For account executives, sales leaders, and BDRs working active deals.

igpt-cs — Customer success intelligence

For CSMs, account managers, and customer success leaders.

igpt-finance — Financial intelligence

For finance teams, controllers, bookkeepers, and anyone managing their own books.

igpt-hr — HR intelligence

For HR business partners, people leaders, and recruiters managing internal talent operations.

igpt-recruiting — Recruiting intelligence

For in-house recruiters, talent acquisition leaders, and external agency recruiters.

igpt-executive — Founder / CEO intelligence

For founders, CEOs, and senior leaders. Email is where the most important decisions, risks, and relationships live for an exec.

igpt-projects — Project & operations intelligence

For project managers, program managers, and operators running cross-functional initiatives.

igpt-marketing — Marketing & partnerships intelligence

For marketing leaders, brand managers, and growth operators.

igpt-procurement — Procurement & vendor intelligence

For procurement teams, category managers, and operators handling vendor relationships.

igpt-it — IT & security intelligence

For IT teams, sysadmins, and IT managers.

igpt-realestate — Real estate agent intelligence

For residential and commercial agents, brokers, and team leaders.

igpt-consulting — Consultant / freelancer intelligence

For independent consultants, boutique agencies, and consulting firms.

igpt-research — Research & academia intelligence

For academic researchers, principal investigators, and research lab leaders.

Architecture

Each plugin in this marketplace bundles four kinds of artifact, all auto-discovered by Claude Code from convention-based folder names.

Skills

A skill is a single focused workflow: collect a few inputs from the user, query iGPT against their email, return a structured answer.

Each skill lives at plugins/<plugin>/skills/<skill-name>/SKILL.md and follows a consistent template:

1. Step 1 — collect user-supplied values. Variables like [time_range] and [scope] are defined with defaults and descriptions. The agent or skill asks the user for any value without a default.
2. Step 2 — query iGPT with structured parameters (search-style filtering by date range, keywords, scope).
3. Step 3 — call the iGPT ask tool with the user's variables substituted into a prompt template, and a strict JSON output schema.

Skills always return schema-validated JSON to the LLM, which is what produces the clean rendered chat response the user sees.

Agents

An agent is a higher-level coordinator for a plugin. Each plugin ships exactly one agent at plugins/<plugin>/agents/<plugin>.md that handles broad, cross-skill questions ("walk me through what needs my attention this week") by routing to the right skills, running them in parallel, and synthesizing a prioritized response.

Agents also enforce domain-specific guardrails:

• Verbatim commercial terms in finance and procurement
• Patterns-not-judgments in HR and recruiting
• Stage discipline + contrarian signals in sales
• Faithful decision reconstruction in projects
• And so on across all 13

Plugins

A plugin is a folder containing one set of skills, one agent, an .mcp.json for MCP server registration, and a plugin.json with metadata. Plugins are auto-discovered by Claude Code from the marketplace manifest.

Marketplace

The marketplace manifest at .claude-plugin/marketplace.json lists all 13 plugins with their categories, keywords, and source paths.

How it works

You ask in chat
        ↓
Agent (or skill) routes the question
        ↓
Skill queries iGPT against your connected email
        ↓
iGPT returns structured findings (with evidence)
        ↓
Claude renders the findings into readable markdown

A plain-language question becomes the right query against your email, returns evidence-grounded structured findings, and gets rendered into a readable answer. You see the rendered answer in chat; integrators can also work with the structured output directly.

Privacy and data flow

Your email stays with iGPT. The plugin doesn't pipe raw email to Claude. You connect your email account to iGPT through OAuth (read-only); iGPT processes the email on its side; only the structured findings — never the raw inbox — reach Claude.

OAuth, not credentials. No API keys, passwords, or access tokens are ever shared with the plugin or pasted into chat. Authentication happens through a standard OAuth flow in the browser.

Read-only. The iGPT MCP only reads from your connected email. The skills are designed for analysis and synthesis — they never send email on your behalf.

Per-user authorization. Each user installs and authorizes their own iGPT connection. Multi-user deployments preserve user-level access controls — no cross-user email leakage.

Sensitive content care. Several plugins enforce additional rules at the skill and agent layer:

• igpt-recruiting: candidate names handled as sensitive data; hiring manager performance described as patterns, not character judgments.
• igpt-executive: measured language for board / fundraising / settlement / compensation contexts; minimized raw quotation.

See each plugin's README.md for the full set of role-specific guardrails.

Repository structure

iGPT-SKILLS-Building-Blocks/
│
├── .claude-plugin/
│   └── marketplace.json
│
├── shared/
│   └── mcp-guard.md
│
├── plugins/
│   │
│   ├── igpt-sales/
│   │   ├── .claude-plugin/
│   │   │   └── plugin.json
│   │   ├── .mcp.json
│   │   ├── README.md
│   │   ├── agents/
│   │   │   └── igpt-sales.md
│   │   └── skills/
│   │       ├── competitor-mentions/SKILL.md
│   │       ├── cross-sell-upsell-signals/SKILL.md
│   │       ├── deal-friction-detector/SKILL.md
│   │       ├── deal-open-items/SKILL.md
│   │       ├── follow-up-radar/SKILL.md
│   │       ├── meeting-prep-briefing/SKILL.md
│   │       └── past-solutions-matcher/SKILL.md
│   │
│   ├── igpt-cs/
│   ├── igpt-finance/
│   ├── igpt-hr/
│   ├── igpt-recruiting/
│   ├── igpt-executive/
│   ├── igpt-projects/
│   ├── igpt-marketing/
│   ├── igpt-procurement/
│   ├── igpt-it/
│   ├── igpt-realestate/
│   ├── igpt-consulting/
│   └── igpt-research/
│       (each plugin follows the same structure as igpt-sales above)
│
├── CLAUDE.md          # Internal repo conventions for contributors
└── README.md          # This file

Each plugin folder contains exactly:

• .claude-plugin/plugin.json — plugin metadata (name, description, version, author, license)
• .mcp.json — MCP server registration (type: "http", URL points to iGPT)
• README.md — buyer-facing front door with chat examples, install instructions, output schemas
• agents/<plugin>.md — the per-plugin agent
• skills/<skill>/SKILL.md — one folder per skill, each with the [var]-driven workflow

Customization and extending

Adjust an existing skill

Each SKILL.md is human-readable and self-contained. Edit the workflow's variable defaults, search keywords, prompt template, or output schema in place. The skill is ready to use as soon as the file is saved — no build step.

Add a new skill

Create a folder under plugins/<plugin>/skills/ with a SKILL.md that follows the project pattern:

• YAML frontmatter with name, description, metadata.version
• A 3-step workflow: collect user inputs (Step 1), query iGPT (Step 2), call ask with input and output_format (Step 3)
• An inline ## Prerequisites block (the standard one used across the marketplace)

Then add the skill to the corresponding agent's "Available skills" list so the agent knows about it.

Add a new plugin

Copy the structure of an existing plugin (e.g. plugins/igpt-consulting/), update plugin.json, write skills and one agent, and add an entry to .claude-plugin/marketplace.json.

Use the prompts and schemas outside MCP

Every SKILL.md documents its prompt template and output schema in plain text. They work directly against the iGPT API without going through MCP — useful for backend integrations, batch processing, or custom UIs. See docs.igpt.ai for direct API docs.

Conventions

This repository follows a small set of strict conventions documented in CLAUDE.md:

• SKILL.md YAML frontmatter — single-line description, nested metadata.version
• .mcp.json shape — {type: "http", url: "https://mcp.igpt.ai/"}, no extra fields
• Inline Prerequisites blocks — self-contained per skill, with the canonical shared/mcp-guard.md as fallback
• Workflow template — [var] placeholders, step-1 user input collection, step-3 ask with input only
• Per-plugin README pattern — practitioner-first voice, "What you can ask", "What you see", "How it works", install paths, collapsible JSON for developers
• Per-plugin agent — one per plugin at agents/<plugin>.md with domain-specific guardrails

Contributors should read CLAUDE.md before making structural changes.

Versioning

Each SKILL.md and agent file carries a metadata.version field (currently 1.0.0 across all 76 skills and 13 agents). Bump the version when behavior changes meaningfully — output schema changes, prompt rewrites, default value shifts.

The marketplace itself follows semantic versioning at the plugin level (see each plugin.json).

Contributing

Contributions are welcome. Please read CLAUDE.md first — it documents the project conventions for skills, agents, READMEs, and .mcp.json configs. PRs that follow those conventions land faster.

A few specific notes:

• Don't hardcode plugin counts in plugin-level READMEs — say "all our role-specific plugins" so they don't go stale.
• Use the [var] placeholder convention in SKILL.md workflows. Hardcoded values inside ask prompts are a no-go.
• Domain-specific guardrails live at the agent body level. Surface them in plugin READMEs' "How it works" sections too.

Related

• iGPT website — sign up for the iGPT account that backs these plugins
• iGPT API docs — direct API reference for backend integrations
• iGPT Python SDK — pip install igptai
• iGPT Node.js SDK — npm install igptai
• Playground — try queries interactively before writing code

License

MIT