Integrations Sh

integrations.sh CLI

The integrations CLI queries integrations.sh, an open registry that answers "what does this service expose to agents, and exactly how do I authenticate to each interface?" for thousands of services.

Run it with npx (no install needed):

npx integrations search stripe

Or install once: npm i -g integrations — the command is integrations.

Output contract: when stdout is piped (or --json is passed), every command emits exactly one JSON document on stdout; diagnostics go to stderr. You are almost always piping — parse the JSON, don't scrape text.

Workflow

1. Search when you have a service name, not a domain:

integrations search stripe --json
integrations search "issue tracking" --kind mcp --json

--kind mcp|openapi|graphql|cli narrows to one surface type, --limit caps results. Results are domain-level: { domain, name, description, kinds[], url }.

2. Look up the domain — the main call. One lookup returns every known surface for a service plus its auth requirements:

integrations stripe.com --json

How to read the response:

• surfaces[] is a discriminated union on type:
  • http — REST API; spec is the OpenAPI spec URL, url the base URL
  • graphql — url is the endpoint, spec optional schema
  • mcp — url is the connect endpoint, plus transports
  • cli — command name and packages[] install options
• credentials is a registry keyed by id; surfaces reference these ids. Each credential has a type (api_key, bearer, oauth2, basic, …), a generateUrl (where to mint it), and setup (markdown acquisition steps).
• Each surface's auth: status: "none" means confirmed public; status: "required" lists entries[] — alternatives (OR) — where each entry's use[] lists credentials needed together (AND); status: "unknown" means not yet determined.

Not found means the domain isn't cataloged yet — escalate to step 3.

3. Detect / discover when the lookup came back empty or stale:

integrations detect acme.com --json      # fast, deterministic probe
integrations discover acme.com --json    # full agentic discovery

detect checks the domain's well-known manifests (integrations.json, MCP server cards, llms.txt, OpenAPI catalogs) and live capabilities. discover runs a doc-reading agent server-side; it can take up to a minute and is rate-limited to 3 requests per 60s — call it once per domain and reuse the result, never in a loop.

Every fact in the registry carries a basis tag: detected (machine-verified signal), discovered (read from docs by an agent), or declared (published by the service owner). Trust declared/detected over discovered when they conflict.

Choosing a surface

When a service exposes several surfaces, prefer in order: an official MCP server (agent-native), an OpenAPI-specced HTTP API (typed, tool-compilable), GraphQL, then CLI.

More commands

• integrations help — the full command list (the CLI grows automatically as the API does; integrations help <cmd> for flags)
• integrations mcp — connection snippets for the hosted MCP server, if you prefer MCP tools over shelling out
• integrations skill — prints this skill (for self-install: integrations skill > ~/.claude/skills/integrations-sh/SKILL.md)
• --no-cache bypasses the cached API description; INTEGRATIONS_BASE points the CLI at another host

Without the CLI

The same data is plain HTTPS (public, no key, CORS *): https://integrations.sh/api/search?q=…, /api/{domain}/surface, /api/{domain}/detect, /api/{domain}/discover, bulk index at /api.json, spec at /openapi.json, MCP server at https://integrations.sh/mcp (streamable-http; tools detect and discover).