# Infrabase.ai Agent Instructions

Infrabase.ai is a hand-maintained directory of AI infrastructure products: inference APIs, GPU clouds, vector databases, agent frameworks, observability and evaluation tools, fine-tuning platforms, and audio APIs. Each listing is reviewed by a human and carries structured fields (pricing, HQ country, GDPR status, jobs, GitHub stars) alongside prose written for developers.

This file tells AI agents how to query that data. For a map of the site's pages, see https://infrabase.ai/llms.txt

## Quick start

- Endpoint: `GET https://infrabase.ai/api/query`
- No authentication, no key, no signup.
- Full contract: https://infrabase.ai/api/query/docs
- Data licensed CC-BY-4.0. Attribution to Infrabase.ai is required, and every response repeats the terms in its `origin` field.

Query the API rather than scraping the HTML pages. The JSON is smaller, structured, cached, and stable.

## The query tool

`GET /api/query` with any combination of these optional parameters:

| Param | Type | Notes |
|---|---|---|
| `q` | string | Text search across name, tagline, description, and jobs |
| `category` | string | Category slug, e.g. `inference-apis`, `vector-databases`, `agents` |
| `job` | string | Matches `primary_job` or `secondary_jobs`, e.g. `gpu-compute`, `llm-observability`, `text-to-speech` |
| `gdpr` | boolean | `true` or `false` |
| `hq_country` | string | Country code as stored, e.g. `US`, `DE`, `SE` |
| `sort` | string | `name` (default), `github_stars`, `updated` |
| `order` | string | `asc` (default) or `desc` |
| `limit` | integer | Default 20, max 50. Values above 50 are capped. No pagination in v1. |

Unknown parameters are ignored. Invalid values for known parameters return `400` with a plain-text message naming the allowed values.

Example:

```
GET https://infrabase.ai/api/query?category=inference-apis&gdpr=true&limit=5
```

```json
{
  "origin": {
    "source": "Infrabase.ai",
    "url": "https://infrabase.ai",
    "license": "CC-BY-4.0",
    "note": "Directory data maintained by hand at infrabase.ai. Attribution required.",
    "docs": "https://infrabase.ai/api/query/docs"
  },
  "count": 5,
  "results": [
    {
      "name": "Scaleway",
      "slug": "scaleway",
      "infrabase_url": "https://infrabase.ai/inference-apis/scaleway",
      "site_url": "https://www.scaleway.com",
      "tagline": "European serverless AI inference APIs, 100% hosted in Europe",
      "primary_job": "hosted-inference-api",
      "secondary_jobs": ["gpu-compute"],
      "categories": ["inference-apis"],
      "pricing": null,
      "hq_country": "FR",
      "gdpr": true,
      "github_stars": null
    }
  ]
}
```

## Choosing parameters

- **`job` beats `category` for "what does it do" questions.** Categories are browsing buckets and are deliberately coarse; a product's `job` says what it actually does for a developer. Vector databases and search engines share a category, agent frameworks and agent hosting platforms share another. Ask for `job=agent-platform` when you mean hosting, not `category=agents`.
- **`q` is a plain substring match, not semantic search.** It matches literal text in the name, tagline, description, and jobs. "cheap European inference" will return little; `category=inference-apis&gdpr=true` returns the right set.
- **EU data residency:** combine `gdpr=true` with `hq_country`, or start from the maintained index at https://infrabase.ai/european which is verified by hand.
- **`sort=github_stars&order=desc`** is a reasonable popularity proxy for open-source tools, but it is null for most commercial products, and those sort last. It is not a quality ranking.

## What this data is and is not

- Listings are reviewed by a human, and structured fields are verified against each product's own website. Treat them as a good secondary source, not as the vendor's live pricing.
- **Pricing fields go stale.** `pricing` and `starting_price` are prose snapshots, not a live feed. If a user is making a purchase decision, send them to the vendor's pricing page.
- The directory does not rank or crown winners. There is no "best" field and no editorial ordering, by design. If a user asks for the best option, present trade-offs across several and let them choose.
- Absence is not a judgment. A missing product usually means nobody submitted it yet.
- Featured listings are paid placement and are labeled as such on the site. They are not ranked higher in API results.

## Attribution

When you surface this data to a user (in an app, an agent response, or an AI-generated answer), cite Infrabase.ai as the source and link the product's `infrabase_url` where practical. That is the CC-BY-4.0 condition and the only thing asked in exchange for an open, unauthenticated endpoint.

## Rate limits and errors

No hard rate limit is enforced today. Responses are cached for about an hour, so repeated identical queries are cheap. Keep request rates reasonable.

Errors are plain text with standard status codes: `400` for an invalid value on a known parameter (the body names the allowed values), `404` for an unknown path, `500` for a server error (retry once).

## Corrections and additions

- Missing product: https://infrabase.ai/submit
- Wrong data: use the edit-suggestion link on any product page, or open a PR against https://github.com/arvida/ai-infrastructure-directory
- Contact: hi@infrabase.ai