How do I use griptape-mcp?

1. Click on "Install Server". 2. Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state. 3. In the chat, type @ followed by the MCP server name and your instructions, e.g., "@griptape-mcp show me how to add conversation memory to a Griptape Agent" That's it! The server will respond to your query, and you can continue using it as needed. Here is a step-by-step guide with screenshots.

griptape-mcp by KianSalem

Your AI keeps hallucinating Griptape method names. You ask it to build an Agent, it writes confident code that doesn't exist. You paste the error. It apologizes and writes different wrong code.

griptape-mcp fixes this. It ships the entire Griptape documentation — 84 framework pages, 125 nodes, 714 real code examples — as a pre-built SQLite database your AI can actually search. No hallucinations. No outdated training data. No API keys.

Install it once. Your AI figures out the rest.

What's inside the box


📚 84 framework pages	Agents, Pipelines, Workflows, Tools, Drivers, Engines, RAG, and more
🧩 125 visual nodes	Every node in Griptape Nodes across 17 categories
🔍 Full-text search	SQLite FTS5 — fast, typo-tolerant, ranked by relevance
💻 714 code examples	Real, working snippets pulled straight from the official docs
📦 Ships ready to go	Pre-built database included. No network calls, no API keys, no scraping at query time.

What it looks like

Once connected, your AI uses the tools automatically. No prompting required.

You: How do I add conversation memory to an Agent?

Claude: (calls search_docs("conversation memory")) Found it. Here's the pattern:

from griptape.structures import Agent
from griptape.memory.structure import ConversationMemory

agent = Agent(conversation_memory=ConversationMemory())

No guessing. No hallucinated imports. Just the actual docs.

Get running in 60 seconds

Step 1 — Install

pip install griptape-mcp

Step 2 — Connect your client

Add to your config file:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json
Linux: ~/.config/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "griptape-docs": {
      "command": "griptape-mcp"
    }
  }
}

Restart Claude Desktop. Look for the 🔨 icon — that means tools are loaded.

claude mcp add griptape-docs griptape-mcp

Done. That's genuinely it.

Or add manually to your MCP settings:

{
  "mcpServers": {
    "griptape-docs": {
      "command": "griptape-mcp"
    }
  }
}

docker build -t griptape-mcp .

{
  "mcpServers": {
    "griptape-docs": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "griptape-mcp"]
    }
  }
}

Step 3 — Ask anything

"How do I create a Griptape Agent with custom tools?"
"What's the difference between a Pipeline and a Workflow?"
"Show me a RAG pipeline example in Griptape"
"What image processing nodes does Griptape Nodes have?"
"Find me code examples for conversation memory"

What your AI can look up

Six tools your AI assistant can call, all read-only and fast:

Tool	What it does
`search_docs`	Full-text search across all Griptape documentation. Filter by `"framework"`, `"nodes"`, or `"all"`.
`get_page`	Pull the complete content of any doc page by URL or title. Includes sections and code blocks.
`search_griptape_nodes`	Find nodes by name, description, or category. Returns descriptions and doc links.
`get_node_details`	Deep dive on a single node — full description, config, and code examples.
`list_categories`	Browse what's available: 8 framework sections + 17 node categories with counts.
`get_code_examples`	Search for working code snippets by topic. Great for "show me how to..." questions.

How it works

The architecture is intentionally boring — a SQLite file shipped inside the pip package, opened read-only at query time. No network calls happen during conversations.

┌──────────────────────────────────────────────────────────┐
│                    At build time                         │
│                                                          │
│  docs.griptape.ai ──┐                                   │
│                      ├──▶ scraper ──▶ SQLite + FTS5      │
│  GitHub markdown ────┘              (shipped in package) │
└──────────────────────────────────────────────────────────┘

┌──────────────────────────────────────────────────────────┐
│                    At query time                         │
│                                                          │
│  Claude / AI ◀──stdio──▶ griptape-mcp ◀──▶ SQLite (ro)  │
│                                                          │
│  No network calls. No API keys. Just a local subprocess. │
└──────────────────────────────────────────────────────────┘

Scrapers crawl docs.griptape.ai (via sitemap) and Griptape Nodes (via GitHub markdown)
Content gets parsed into structured pieces: titles, headings, code blocks, node metadata
Everything lands in a SQLite database with FTS5 full-text search indexes
That database ships inside the pip package — nothing to fetch at runtime
The MCP server opens it read-only and exposes 6 search/lookup tools over stdio
A nightly GitHub Actions job re-scrapes and rebuilds to stay current

Node coverage

125 nodes across 17 categories:

Category	Count	Category	Count
Image	32	Lists	17
Video	18	Text	12
Config	9	Number	7
Tools	7	JSON	5
Dict	4	Audio	3
Execution	3	Rules	2
Adv. Media Library	2	Agents	1
Convert	1	Utils	1
3D	1

Development

Setup

git clone https://github.com/KianSalem/griptape-mcp.git
cd griptape-mcp
pip install -e ".[dev]"

Rebuild the docs database

cd scripts
python build_db.py

This scrapes both documentation sources and writes to src/griptape_mcp/data/griptape.db. If the website rate-limits you, the build script automatically falls back to scraping GitHub markdown.

Run against a local database

GRIPTAPE_MCP_DB_PATH=./griptape.db griptape-mcp

Validate

python scripts/validate_db.py src/griptape_mcp/data/griptape.db

  [PASS] Framework pages > 10 - got 84
  [PASS] Nodes pages > 10 - got 164
  [PASS] Nodes extracted > 20 - got 125
  [PASS] Sections > 50 - got 2263
  [PASS] Code examples > 10 - got 714
  [PASS] FTS search works - 'agent' matched 79 pages
  [PASS] Multiple node categories - got 17
  ALL CHECKS PASSED

Project structure

griptape-mcp/
├── src/griptape_mcp/
│   ├── server.py               ← MCP tools (FastMCP)
│   ├── db.py                   ← Schema, queries, FTS
│   ├── __main__.py             ← Entry point
│   └── data/griptape.db        ← Pre-built database (14 MB)
├── scripts/
│   ├── build_db.py             ← Orchestrates full rebuild
│   ├── scrape_framework.py     ← Crawls docs.griptape.ai
│   ├── scrape_nodes.py         ← Crawls docs.griptapenodes.com
│   ├── scrape_nodes_github.py  ← GitHub fallback scraper
│   └── validate_db.py          ← Post-build validation
├── .github/workflows/
│   ├── rebuild-db.yml          ← Nightly CI rebuild
│   └── publish.yml             ← PyPI publish on release
├── Dockerfile
└── pyproject.toml

Contributing

See CONTRIBUTING.md.

License

MIT — do whatever you want with it.

griptape-mcp