Promethist for Claude

Files, Links and Conversation Data

An MCP conversation carries text, not files. Where files are involved, the server hands out links instead.

Uploading Files

Knowledge documents and the images and videos of interactive content are uploaded through a one-time link:

Claude calls upload_knowledge_file or upload_interactive_content_file. The server checks that you may edit the project and returns a ready-made curl command with a placeholder for the path of your file.
The command is run on the computer where the file is. Claude Code or Cowork can run it for you; with Claude on the web, run it yourself in a terminal after replacing the placeholder.
The file is processed when it arrives — a knowledge document is read and indexed — and Claude can check the result afterwards.

Upload	Size limit	Formats
Knowledge document	10 MB	PDF, Word, PowerPoint, Excel, HTML, CSV, plain text and images (JPEG, PNG, TIFF)
Image of interactive content	25 MB	PNG or JPEG
Video of interactive content	500 MB	MP4, MOV or WebM

Upload

Size limit

Formats

Knowledge document

10 MB

PDF, Word, PowerPoint, Excel, HTML, CSV, plain text and images (JPEG, PNG, TIFF)

Image of interactive content

25 MB

PNG or JPEG

Video of interactive content

500 MB

MP4, MOV or WebM

Right after the upload the element uses the video as it was uploaded. Shortly afterwards the video is converted to an MP4 that plays on phones and in browsers alike — H.264 video of at most 1920 pixels and 3 Mbit/s with AAC audio — and the element shows the converted video from then on; the uploaded file is kept. The same happens to a video uploaded in the Studio. A video that cannot be converted — larger than 4K, or in an unusual format — stays as it was uploaded, and so does a video given by its web address.

An upload link works once and expires after one hour. Treat it like a password until it is used: whoever holds it can upload the file.

Logos cannot be uploaded through Claude; set a logo URL instead, or upload the logo in the Studio.

Downloading Exports

export_analytics (as Excel or CSV) and export_conversation_data return a download link instead of the data itself. The link is valid for 24 hours and anyone who has it can download the file, without signing in. Share it only as you would share the exported data.

Export	Content
Analytics report	A summary of the time range, one row per conversation (user, agent, start and end, turns, duration and — if your account allows — the transcript) and the evaluation results — if your account allows. As JSON it is returned directly; as Excel or CSV as a link. A report covers at most 100,000 conversations; a time range with more is refused with a request to narrow it. An Excel report is also refused when its evaluation results do not fit on one sheet; export CSV or JSON instead. An Excel cell holds at most 32,767 characters, so a longer text, such as a very long transcript, is cut there with a note; JSON and CSV keep it whole.
Conversation data	Raw records of conversations — turns, or the model calls behind them — as NDJSON (one record per line) or, for turns, CSV. Up to 100,000 records per export. An NDJSON export can be narrowed to named fields, which is what keeps an export of model calls — whole records run to hundreds of kilobytes each — to one file; the identifiers and the timestamp are always kept, so a narrowed record can still be matched to its conversation.

Export

Content

Analytics report

A summary of the time range, one row per conversation (user, agent, start and end, turns, duration and — if your account allows — the transcript) and the evaluation results — if your account allows. As JSON it is returned directly; as Excel or CSV as a link. A report covers at most 100,000 conversations; a time range with more is refused with a request to narrow it. An Excel report is also refused when its evaluation results do not fit on one sheet; export CSV or JSON instead. An Excel cell holds at most 32,767 characters, so a longer text, such as a very long transcript, is cut there with a note; JSON and CSV keep it whole.

Conversation data

Raw records of conversations — turns, or the model calls behind them — as NDJSON (one record per line) or, for turns, CSV. Up to 100,000 records per export. An NDJSON export can be narrowed to named fields, which is what keeps an export of model calls — whole records run to hundreds of kilobytes each — to one file; the identifiers and the timestamp are always kept, so a narrowed record can still be matched to its conversation.

What Conversation Data Claude Can Read

Tools Who can use them

Tools	Who can use them
`list_sessions`, `get_session_transcript`	Anyone with access to the project. Like the Studio’s conversation list, they leave out test conversations and conversations with fewer than two turns.
`search_turns`, `search_llm_logs`, `describe_conversation_data`, `export_conversation_data`	Members with a role on the account; a membership in a single project is not enough. These are the tools that also reach test conversations.

list_sessions, get_session_transcript

Anyone with access to the project. Like the Studio’s conversation list, they leave out test conversations and conversations with fewer than two turns.

search_turns, search_llm_logs, describe_conversation_data, export_conversation_data

Members with a role on the account; a membership in a single project is not enough. These are the tools that also reach test conversations.

All of them respect your account’s Display full conversation data setting: when it is off, transcripts and raw conversation data are not available through Claude.

Links to the Studio

When a result refers to an account, project or agent, it carries the address of its page in the Studio (webUrl), so you can open what Claude created or changed directly.