# The AI Inspector

> The AI Inspector captures the literal JSON request that left your browser: how to open it, what each frame proves, which requests carry a body, and what it leaves out.

- Canonical: https://big-agi.com/docs/feature-ai-inspector
- Minimum tier: open  [Open ⊂ Free ⊂ Pro - a tier only ever ADDS to the Open baseline]
- Kind: reference · Last content update generation: 2026-07-29 · Review by: 2026-10-25
- Answers: Can I see the literal request that left my browser? · What exactly did the model receive? · How do I keep a record of a run? · Why is the request body missing from a capture?

The AI Inspector is an **expert-level developer tool** that shows the literal JSON request that left your browser for the turn you just sent - proof of what was sent, and of which model and settings produced it.

The prompts Big-AGI assembles are open to you: the exact bytes that leave for the AI service. It is the same packet a browser's network view would show, inside the app and without DevTools. Open, Free and Pro all carry it.

On first open the dialog is armed and empty - the request picker reads "No requests available", and the panel waits for your next AI request.

![The AI Request Inspector freshly opened: the capture toggle on, a Select Request picker reading No requests available, History Size set to Keep 10 requests, a Clear History button, and the empty state Ready to capture - Your next AI request will be captured here](/docs/feature-ai-inspector/inspector-empty-bonobo.png)

## Open the inspector

| How to open it | What it says |
|---|---|
| Settings -> Tools -> AI Inspector | Inspect live AI requests |
| `Ctrl + Shift + A` | the same dialog |

The dialog shows each captured frame with its transport chip, the decoded URL, and the pretty-printed body.

![The AI Request Inspector with a captured conversation frame: Edge Server and Done chips, the decoded URL, and the pretty-printed JSON body with the model, tools and thinking fields](/docs/feature-ai-inspector/inspector-frame.png)

That request body holds the assembled system prompt and the history as it was sent. The model, its parameters, the tool declarations and each attachment part are fields of it too. The first open in a session arms capture by itself; after that the switch in the title bar controls it. Turn it on, send a message, and the frame appears. History holds 5, 10, 20 or 50 frames, 10 by default, in memory only and with no export button.

## What each frame shows

| Field | What it is |
|---|---|
| Request number | a counter for this session |
| Status | two chips: the transport - **Direct Connection** or **Edge Server** - and **Done** or **In Progress** |
| Date | local timestamp |
| **URL** | the upstream endpoint, decoded |
| Context | what asked for it: `conversation`, `beam-scatter`, `chat-ai-title`, and so on |
| Reference | the id of the thing that triggered it |
| Headers | the request headers block |
| **Body** | the literal JSON request, pretty-printed, with a byte-size readout |
| Particles | the response stream as discrete timestamped events; aborted ones flagged |

The transport chip records which path the request took. For a record of a run - a methods section, a file note - copy the **Body** block, which copies as JSON. Take the header line with it: request number, transport chip, date, URL, context, reference. Make the copy before you reload.

> **Warning:**
> Read the URL line before you copy or screenshot a capture: it is displayed decoded, and on the browser route it can carry a key. [Keep your key safe](/docs/key-storage).

## Which requests carry a body

| The frame | Contexts |
|---|---|
| Body present, in a production build | `conversation` · `scratch-chat` · `chat-react-turn` · `beam-scatter` · `beam-gather` · `beam-followup` |
| Body absent, response particles only | `chat-ai-title` · `chat-attachment-prompts` · `aifn-image-caption` · `summarize` · the follow-up diagram and chart generators |

Your own turns and your Beam calls are the first list. The background utility calls are the second, and they are skipped by the frame picker's auto-selection. To audit those payloads, use the browser's own network view, below. An instance running a development build shows headers, the profiler and every context.

## Advanced

The footer expands into a JSON request injection field and a Disable Streaming toggle.

![The dialog footer expanded: JSON request injection with Apply and Clear, and a Disable Streaming toggle reading Force all AI requests to not stream](/docs/feature-ai-inspector/advanced.png)

Two controls for advanced users; we discourage everyone else from touching them. The **AI Injector** merges a JSON blob into every request that goes out, overriding fields of the body - a capture taken while it is armed is not a record of normal behaviour, so leave it alone when you are auditing. **Disable Streaming** forces every AI request to run non-streaming.

## What the inspector leaves out

- Request headers are blanked in every production build, on both transports, so a capture never shows an API key.
- Response headers, the raw stream bytes and the transport envelope - particles are the parsed stream, not the wire.
- Request bodies for the background contexts above, in production builds.
- Non-model traffic entirely: browsing, search, image generation, account and analytics calls. This is an *AI request* inspector.
- Anything the AI service does after receipt.

For a request that failed rather than one you did not expect: [Error messages](/docs/issue-messages).

## Related

- [What leaves the browser](/docs/privacy-data-flow)
- [What Direct Connection does](/docs/feature-direct-connection)
- [Keep your key safe](/docs/key-storage)
- [Error messages](/docs/issue-messages)