# Agent skill

> A SightRadar skill for Claude Code, OpenClaw and Codex: the API, the CLI, costs and consent rules, loaded only when a task needs them.

Source: https://sightradar.com/docs/agent-skill

> **Coming soon.** The skill is not published for download yet. This page describes what it contains and where it goes, so you can install it the day it is out.

A skill is a folder with a `SKILL.md` file that an agent loads on demand. The agent reads the one-line description at the start of a session, and only when a task matches does it read the rest. Asking to "find this guest's photos in the gallery" or "move our IndexFaces calls off Rekognition" pulls in the SightRadar instructions; unrelated work never sees them.

The SightRadar skill is about 170 lines of instructions and three reference files. It covers the REST API and the [CLI](/docs/sdks/cli), so the agent can work through `curl` alone or through `sightradar` when it is installed.

## What it covers

| Area       | What the agent learns                                                                                      |
| ---------- | ---------------------------------------------------------------------------------------------------------- |
| Auth       | Read the key from `$SIGHTRADAR_API_KEY`, or `sightradar login` from stdin. Never print it                  |
| Cost       | The credit price of every operation, and that billing is per photo, not per face                           |
| Images     | URL, GCS key, multipart file or raw bytes; which endpoints accept which                                    |
| Operations | Collections, index, batch, search by image, search by id, selfies, compare, detect, delete, balance, usage |
| Thresholds | Cosine similarity 0 to 1, calibrate on your own photos, accept / review / reject bands                     |
| Errors     | What each status means, what to do next, and the matching CLI exit code                                    |
| References | Full response fields, batch webhooks and signature checks, and the Rekognition mapping                     |

Each operation shows the `curl` call and the equivalent CLI command side by side.

## Rules the agent follows

The skill opens with four rules, ahead of any API detail:

* **Billable calls spend your money.** The agent confirms before loops or batches, previews with `--dry-run`, checks `sightradar balance` first, and sends an `Idempotency-Key` on any call it might retry.
* **The key stays a variable.** It references `$SIGHTRADAR_API_KEY` and never prints the value, writes it to a file, or puts it in client-side code.
* **Deletes are soft unless you ask.** The agent uses `--immediate` or `--compliance` only when you ask for permanent erasure.
* **A selfie is biometric data.** The agent registers selfies only with the person's consent and for the purpose they expect.

These match the [Acceptable Use Policy](/aup). SightRadar is for consented use, such as guests finding their own event photos or 1:1 verification a person has agreed to. The skill does not help with identifying people who have not opted in.

## Install

Copy the skill folder (the one containing `SKILL.md`, named `sightradar`) into the skills directory of your agent.

_Claude Code_
```bash
# For every project on this machine
cp -r sightradar ~/.claude/skills/sightradar

# Or for one repository, committed so the team gets it
cp -r sightradar .claude/skills/sightradar
```

Claude Code lists it under `/skills` and loads it when a task matches. You can also invoke it directly with `/sightradar`.

_OpenClaw_
```bash
# This workspace
cp -r sightradar <workspace>/skills/sightradar

# All workspaces
cp -r sightradar ~/.openclaw/skills/sightradar
```

Set `SIGHTRADAR_API_KEY` in the environment OpenClaw runs in.

_Codex_
```bash
# For your user
cp -r sightradar ~/.agents/skills/sightradar

# Or for one repository
cp -r sightradar .agents/skills/sightradar
```

Codex scans `.agents/skills` from the working directory up to the repository root.

Other agents that read `SKILL.md` folders work the same way. If yours does not support skills, use [agent setup](/docs/agent-setup) instead.

Then give the agent a key without pasting it into the chat:

```bash
export SIGHTRADAR_API_KEY="frs_..."      # or: printenv SR_KEY | sightradar login
```

## Skill, llms.txt or both

They do different jobs, and they work together.

|                 | Agent skill                                                       | [llms.txt](/docs/agent-setup)                                      |
| --------------- | ----------------------------------------------------------------- | ------------------------------------------------------------------ |
| What it is      | Instructions the agent loads when a task matches                  | A map of every docs page, with a Markdown twin for each            |
| Where it lives  | In your agent's skills folder                                     | `https://sightradar.com/llms.txt`, fetched on demand               |
| Tells the agent | How to act: cost checks, consent, retries, which command to run   | What exists and where the full reference is                        |
| Best for        | Doing the work: indexing a gallery, running a search, a migration | Answering questions and reading the exact contract of one endpoint |

The skill is deliberately short and points at `llms.txt` and the `.md` pages for anything it does not inline, so the agent fetches the full reference only when it needs it. If you install the skill, you do not also need the `AGENTS.md` block from [agent setup](/docs/agent-setup), but it does no harm.

An MCP server is coming soon as a separate install.