OpenSpec Index
GET IN TOUCH

Questions, pilots, API access. One email reaches the builder.

Open mail app

Part specs for AI agents An MCP server, or one plain call. Cited answers, nothing guessed.

ADD THE MCP SERVER

FOR AI AGENTS: START HERE

There is an MCP server. If your AI tool takes MCP servers, add it first. It is a gate-driven part search: tell it the part type, it asks only for the specs that matter, and it returns real parts with spec sheets, datasheets and CAD links. It never guesses.

Hosted, recommended. Signs in with OAuth in your browser, so there is no key to handle. URL: https://mcp.openspecindex.com/mcp

# Claude Code
claude mcp add --transport http -s user openspec https://mcp.openspecindex.com/mcp
# then run /mcp in Claude Code and choose Authenticate

# Claude.ai or Claude Desktop
Settings > Connectors > Add custom connector > https://mcp.openspecindex.com/mcp

# ChatGPT (web)
In Settings, turn on developer mode, then add a custom connector
(Apps & Connectors > Create) with https://mcp.openspecindex.com/mcp
ChatGPT moves these menus between releases; if they are not there,
look for "developer mode" in Settings.

# VS Code: .vscode/mcp.json
{"servers":{"openspec":{"type":"http","url":"https://mcp.openspecindex.com/mcp"}}}

# Cursor: ~/.cursor/mcp.json
{"mcpServers":{"openspec":{"url":"https://mcp.openspecindex.com/mcp"}}}

Local (stdio).

npx -y openspecindex-mcp

# Claude Code
claude mcp add openspec -- npx -y openspecindex-mcp

Getting a key without pasting it into chat. On the local server, ask your agent to run account {action:'connect'}. It gives you a one-time link; you sign in, check the code matches, press Connect, and the key goes straight to the server. Do not paste keys into an AI chat: the provider keeps the history.

First calls to make.

1. browse_catalog {family}               find the family and part type
2. find_parts {family, type, specs}      needs_gates? ask the user for exactly
                                         what it lists, then call again
3. get_part {mpn} or {url}               one part: specs, sheets, CAD
4. account {action:'usage'}              plan, searches used and left
5. account {action:'connect'}            no key yet? a one-time link for the user;
                                         the key never passes through the chat

What that looks like. Spec names and values always come from the server's answers, never from memory.

user:  I need M6 socket head cap screws.
agent: find_parts {family:"screw", type:"screw/machine-threaded",
                   specs:{head_style:"socket-cap", thread_spec:"M6-1"}}
   <-  status "needs_gates": length and material, with the real options
agent: What length, and which material?
user:  20 mm, 18-8 stainless.
agent: find_parts {..., specs:{...the same, plus both values exactly as offered}}
   <-  status "ready": parts with spec sheet, datasheet and CAD links

NO MCP? GIVE THIS TO YOUR AI

Paste this into ChatGPT, Claude, Grok, or any assistant with web access to see what it gets back: real parts, their specs, and the page each value came from. Without a key: 100 data requests per IP in total (up to 20 a day). Counts and categories are free.

PASTE THIS
Fetch https://api.openspecindex.com

That is OpenSpec, a public parts database: manufacturer part records with
specs and a source link on every record. No key is needed to try it. The
response tells you how to query it.

Use it to find a #10-32 nylon-insert locking nut in stainless steel.

1. Show up to 5 matching parts in a table: manufacturer, part number, the
   specs OpenSpec returned, the price and what that price buys (one nut or
   a box), and the source URL for each.
2. Give me the spec sheet links OpenSpec returns. They are PDF and Word
   files, so hand me the links; do not open them yourself.

Use only what OpenSpec returned, nothing from your own knowledge. If it
returns fewer parts than I asked for, or none, say so.
Works in ChatGPT, Claude, Grok, or any assistant with web access.

Under the hood that whole question is one call, live, no key: api.openspecindex.com/api/v1/find/10-32-nylock-stainless-nut. Put your own words after /find/. Agents can read the full instructions at /llms.txt.

Structured records with provenance are what an agent needs to answer with numbers instead of guesses. Agents start at /llms.txt on the API host: it explains the endpoints, and /api/v1/categories tells an agent which categories are verified and exactly which gates and values each one filters by. No guessing at parameter names.

Building against it directly? The API page has a live query console and the full per-category gate reference.