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

# MCP server overview

> Connect AI clients to Inspack through the Model Context Protocol.

The Inspack MCP server lets any MCP-aware client — Claude Desktop, Cursor, ChatGPT, and others — call Inspack on your behalf. You authenticate once with your existing API key, and your AI assistant can run content reviews directly from chat.

## What you can do

* **Run reviews** from a file drop, a URL, or pasted text — without leaving the AI client.
* **Pass text and image URLs** in a single call; results map back to the spans they apply to.
* **Stay in your workflow** — no glue code, no REST wiring.

## How a review works

When you ask your AI client to "review this," it orchestrates Inspack tools in sequence. You don't call them by hand — the server's instructions tell the model what to do.

1. **`prepare_upload`** — initialize with `readable: {filename, content_type, bytes}` and optional originals. PUT each file from the client's shell to its returned slot URL using every `required_headers` entry verbatim, with no bearer credentials. Backend limits and expiry are authoritative.
2. **`process_upload`** — process with `{asset_id, version}` and an optional `ctx`: omit it for a review upload, or use `inspector_extraction` (with `inspector_type` and optional `on_question`) or `reference_asset`. Follow the version returned when deduplicated. If the result or a later status is `awaiting_input`, show `question.message` and `question.options[].label`, then answer with `ctx: {operation: "answer_question", run_id, question_id, option_id}`. `start_upload` is retired and returns migration guidance.
3. **`get_upload_status`** — read one snapshot and its latest content envelope. Clients choose polling intervals; inspect stages for partial or failed processing. Convert payload text/grounding into text review objects and use `status.latest.render.url` for a visual image object. MCP `review` temporarily fetches canonical protected upload images server-side and converts them to data URLs.
4. **`get_inspectors`** — select inspectors available to your organization.
5. **`review`** — review one shared objects list with `inspector_ids`; returns inspector results and aggregated summaries together.

Public raster URLs and raw text can go directly to review. Download remote documents and use the upload lifecycle. Upload calls consume no review credits. See the [upload contract and migration guide](https://docs.inspack.com/public-upload-api.md) for file shapes, exact headers, authenticated artifact reads, and system-token scoping.

`doc_parse`, `inspect`, `inspect_multiple`, and `aggregate` remain discoverable but are retired. Calls return `tool_retired` migration errors and consume no review credits. See the [review contract](https://docs.inspack.com/public-review-api) for review operations.

## What's needed

* An Inspack API key (see [Authentication](https://docs.inspack.com/authentication.md)).
* An MCP-aware client.

## Next

> **[Connect a client](https://docs.inspack.com/mcp/connect.md)**
>
> Step-by-step setup for Claude Desktop, Cursor, and ChatGPT.

> **[Available tools](https://docs.inspack.com/mcp/tools.md)**
>
> Tools the server exposes and what they return.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.