> ## Documentation Index
> Fetch the complete documentation index at: https://docs.pureframe.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Agent Vision

> Give any AI agent the ability to see and search your video library.

Agent Vision lets any AI model search your video library, retrieve video details, and see matched frames directly — without fetching URLs or managing storage.

## How it works

When an agent calls `search_videos`, results include `thumbnail_base64` — a base64-encoded JPEG of the matched frame. Vision-capable models can see this image in the same turn:

1. Agent searches for a moment in video
2. Pureframe AI returns the matched frame as base64
3. Agent reads the visual content and answers follow-up questions

No URL fetching, no extra round trips.

## Two ways to connect

**MCP** — the simplest path for Claude, Cursor, VS Code, Codex, OpenCode, and any other MCP-compatible client. See [MCP Overview](/agents/mcp-overview) and [Connect an Agent](/agents/connect-an-agent).

**Direct HTTP / function calling** — for OpenAI, Gemini, or any LLM with function calling, fetch the OpenAI-compatible schema and wire it directly:

```python theme={null}
import httpx

BASE = "https://api.pureframe.ai"
HEADERS = {"Authorization": "Bearer pf_..."}

# Get the tool schema (OpenAI format)
schema = httpx.get(f"{BASE}/v1/agent/schema.json", headers=HEADERS).json()

# Call a tool
result = httpx.post(f"{BASE}/v1/agent/call", headers=HEADERS, json={
    "tool": "search_videos",
    "input": {
        "query": "presenter pointing at a chart",
        "collection_id": "col_abc123",
        "limit": 5
    }
}).json()
```

Pass `schema` directly to `client.chat.completions.create(tools=schema)` for OpenAI or the equivalent for other providers.

## Example: search + vision with Claude

```python theme={null}
import anthropic, httpx

client = anthropic.Anthropic()
pf = {"Authorization": "Bearer pf_..."}

# Search for a moment
clips = httpx.post("https://api.pureframe.ai/v1/agent/call",
    headers=pf,
    json={"tool": "search_videos", "input": {"query": "whiteboard diagram", "limit": 1}}
).json()["data"]

# Pass the matched frame directly to Claude
frame_b64 = clips[0]["thumbnail_base64"]

response = client.messages.create(
    model="claude-opus-4-7",
    max_tokens=1024,
    messages=[{
        "role": "user",
        "content": [
            {"type": "image", "source": {"type": "base64", "media_type": "image/jpeg", "data": frame_b64}},
            {"type": "text", "text": "What's written on the whiteboard?"}
        ]
    }]
)
print(response.content[0].text)
```

## Tool response fields

`search_videos` returns a list of clips. Each clip:

| Field                     | Description                                        |
| ------------------------- | -------------------------------------------------- |
| `video_id`                | ID of the source video                             |
| `filename`                | Original filename                                  |
| `start_secs` / `end_secs` | Clip boundaries in seconds                         |
| `relevance_score`         | Relevance from 0 to 1                              |
| `text_snippet`            | Transcribed speech in this clip, if available      |
| `clip_url`                | Presigned URL to stream the video (valid \~1 hour) |
| `thumbnail_url`           | Presigned URL to the matched frame                 |
| `thumbnail_base64`        | Base64 JPEG for direct use with vision models      |

Every answer an agent gives from these fields is traceable back to an exact clip — see [Source-backed Results](/agents/source-backed-results).