AI agents for data and research: where the tokens really go
Researchers, data scientists, bioinformaticians and UX researchers feed their AI agents files, not prompts. That is where the context and the usage limit go.
Who it is for
Researchers & PhD students
For researchers and PhD students who run literature reviews with an AI agent: papers become compact text before they reach the model, so the session and the quota last longer.
Data scientists
For data scientists and analytics engineers who work with an AI agent: notebooks, CSVs and query output reach the model compacted, with the complete data kept on your computer.
In data and research work, an AI agent spends most of its tokens reading, not writing. A literature review feeds it dozens of PDFs, a notebook session feeds it JSON and result sets, a pipeline feeds it logs, and an interview study feeds it hours of transcripts. Tokenade makes those inputs smaller on your computer before the model sees them.
Why these fields adopted agents
The work is file-heavy and repetitive, which is exactly what an agent running in a folder does well. The agent reads the material, runs the scripts and drafts the output; the person checks it.
Anthropic markets Claude Code for genomic analysis in its Claude for Life Sciences announcement (2025-10-20). Researchers publish systematic-review workflows built on it (2026-05-14), and a UX research guide (2026-04-27) describes it coding forty transcripts in an afternoon.
Where the tokens go
The common thread is that the expensive part is the material, not the conversation.
- Documents. Anthropic's PDF documentation counts 1,500 to 3,000 tokens per page for the text, with the page image billed on top (read 2026-10-15).
- Structured data. Notebooks are JSON with embedded outputs; CSVs and query results break into many tokens per cell.
- Logs. Pipeline runs, test suites and training jobs print long, repetitive output that stays in the conversation.
- Media. Interviews and talks arrive as audio or video that has to become text first.
Every one of these stays in the transcript once read, and is re-sent to the model on each following turn. On a Claude subscription, the 5-hour window and the weekly cap are shared across chat, Claude Code and Cowork (claudelimit.com, 2026-09-01), so a heavy reading session also eats into everything else you do with Claude.
What the audiences below have in common
They need exact content back: a verbatim quote, a precise error line, the value in row 4,812. Cutting context blindly saves tokens and loses that. Tokenade folds instead of deleting: documents become text, long outputs become a preview plus a hash, and the full version stays in a local stash the agent can search or restore.
They also handle material that is not theirs to share freely: unpublished manuscripts, patient-derived data, participant interviews. Compaction runs on the computer where you work and nothing passes through Tokenade's servers, but the agent's provider still receives what the agent reads. Tokenade masks credentials; it does not hide people's names. The privacy page lists every field it sends.
Tokenade works with AI agents that run on your computer, such as Claude Code (including the Code tab of Claude Desktop), Codex and Cursor. It does not act in Claude Cowork, and in the Claude Desktop chat only MCP results are optimised.
On long sessions, Tokenade is 39% cheaper than working without it, measured on the THOL benchmark. The details of document, audio and video reading are in Read documents and media.