Nimble Web Search Skills & Plugin

imageLicense: MIT imageVersion

Unlock the web for your AI agents — search, scrape, extract structured data, and run business intelligence workflows, all powered by Nimble's web data infrastructure. One plugin for Claude Code, Cursor, and any platform that supports the Agent Skills spec.

Agent credential protocol: see AUTH.md — the file agents read to authenticate to the Nimble API. Served live at https://nimbleway.com/auth.md.

Skills

Business Research, Marketing, Productivity, and Healthcare skills are one-command workflows. They spawn parallel sub-agents, gather live web data via Nimble APIs, synthesize findings, and deliver structured reports with dates and source URLs. They learn from previous runs and only surface what's new.

Every skill lives directly under skills/. The category above is recorded as metadata.category in each skill's SKILL.md frontmatter — the plugin platforms require a flat skills tree, so categories are metadata rather than folders. See CLAUDE.md.

Web Data Toolkit skills expose Nimble's raw capabilities for any web task and power the business skills under the hood — search and extract for raw data, Extraction Templates for reusable site-specific scraping, and Web Search Agents for open-ended research, enrichment, and dataset building.

Quick Start

1. Install the Nimble CLI

npm i -g @nimble-way/nimble-cli

2. Set your API key

Sign up and grab your key from Account Settings > API Keys.

export NIMBLE_API_KEY="your-api-key-here"

Or add it permanently to ~/.claude/settings.json:

{ "env": { "NIMBLE_API_KEY": "your-api-key-here" } }

3. Add the skills

Any Claude product (Claude Code, Claude Cowork, claude.ai) — recommended:

/plugin install nimble

One command. The plugin's .mcp.json auto-registers as a Connector pointing at the Nimble MCP server over native HTTP with OAuth — no API key header to manage. On first use, run /mcp and authenticate nimble in your browser.

In claude.ai / Cowork the connector appears under Customize → Connectors as Nimble — click Connect and complete the browser login to activate it (you can create a Nimble account inline if you don't have one).

Cursor:

/add-plugin nimble

Or clone the repo and open it in Cursor — the plugin system auto-discovers skills from .cursor-plugin/plugin.json.

Manual install (Codex CLI, raw MCP clients, or when you'd rather use an API key):

claude mcp add --transport http nimble https://mcp.nimbleway.com/mcp \
  --header "Authorization: Bearer ${NIMBLE_API_KEY}"

Restart Claude Code after running this — MCP servers added mid-session aren't available until the next launch.

npx skills CLI:

npx skills add Nimbleway/agent-skills

To install a single skill:

npx skills add Nimbleway/agent-skills --skill competitor-intel

To list available skills:

npx skills add Nimbleway/agent-skills --list

4. Try it

# Business intelligence — ask a question, get a sourced report
"What are my competitors doing this week?"
"Prepare me for my meeting with Jane from Acme Corp"

# Web data toolkit — get data from any website
"Scrape the pricing page at example.com"
nimble search --query "AI agent frameworks" --max-results 10

How It Works

Nimble Web Search Skills follow a shared pattern: preflight (check CLI, load profile) → parallel research (spawn sub-agents for concurrent data gathering) → analysis (synthesize findings, deduplicate against previous runs) → report (structured output with sources) → distribute (offer Notion/Slack delivery).

Core skills expose the Nimble CLI directly — search, extract, map, crawl, run Extraction Templates, and run Web Search Agents.

Local Web Knowledge Wiki

Skills maintain a local web knowledge wiki at ~/.nimble/memory/ — live web intelligence that compounds across sessions, so your agent never starts from scratch. Inspired by Karpathy's LLM Wiki: an AI-maintained, human-owned knowledge base that gets smarter with every run.

• Local-first indexes — per-directory entity catalogs for instant lookup, no vector DB needed
• Obsidian-compatible — [[wiki links]] cross-reference people, companies, and competitors. Open ~/.nimble/memory/ in Obsidian to browse your intelligence graph
• Cross-entity synthesis — patterns across competitors, pricing trends, and market signals are surfaced automatically
• Ad-hoc insights — say "remember this" mid-conversation and it compounds into the right entity page
• Activity log — grep-friendly record of what was learned, when, and by which skill

Every finding carries a verified event date and source URL. Stale signals are dropped, not reported — your context is always current.

Platform Compatibility

All platforms read the same skills/ directory. Platform-specific files coexist without interference.

The repository is also an Agent Plugins v1.0.0 package, the vendor-neutral format used by VS Code, Cursor, GitHub Copilot, Kiro, and ChatGPT & Codex. Root mcp.json is that format's canonical MCP config. Cursor is listed above as reading the user's own .cursor/mcp.json because that is where a Cursor user pastes the snippet — Cursor does not load this repository's root mcp.json.

Why there are two MCP config files

mcp.json and .mcp.json describe the same endpoint in different vocabularies, and no single file satisfies every consumer:

The portable schema sets additionalProperties: false at the root and on every server, so there is no room to carry a second vocabulary in one file. .mcp.json is the file Claude Code auto-registers as a Connector, so the two are kept separate deliberately rather than merged. python3 scripts/check-plugin-manifests.py validates both.

Every skill directory is an immediate child of skills/, with its vertical recorded as metadata.category in the frontmatter. Reference documents inside a skill's references/ directory are named reference.md — never SKILL.md, which would register them as skills on platforms that discover recursively. bash scripts/check-plugin-structure.sh enforces both rules, and python3 scripts/check-plugin-manifests.py validates the per-platform manifests.

Grok Build is the one platform whose MCP server is declared inline in its manifest rather than by a path to .mcp.json. That keeps .mcp.json identical for Claude Code, whose plugin install auto-registers it as a Connector over native HTTP with OAuth. CI asserts the two declarations never drift apart.

CLI Commands

Contributing

Have a web data workflow that should be one command? We'd love new skills — whether it's sales prospecting, lead enrichment, or anything that turns web data into action. See CONTRIBUTING.md for how to build and publish your own.

Links

• Nimble: nimbleway.com
• Documentation: docs.nimbleway.com
• Issues: github.com/Nimbleway/agent-skills/issues
• Email: support@nimbleway.com

License

MIT License — see LICENSE for details.

The MIT license covers the code, skills, and documentation in this repository. It does not grant rights to Nimble's trademarks or brand assets: the Nimble name, the Nimble logo, and the brand marks under assets/ remain the property of Nimble and are not licensed for reuse.