Document work with an AI agent: finance, legal, consulting and marketing
For people who work their documents with an AI agent on their own computer. Tokenade turns PDFs, Office files, exports and web pages into compact text before the agent reads them.
Who it is for
Finance & accounting
For analysts who run Claude Code, Codex or Cursor on filings, PDFs and workbooks: the agent reads compact text instead of raw pages and rows, so the limit lasts longer.
Lawyers & legal teams
For lawyers and legal teams who run Claude Code, Codex or Cursor on a matter folder: the agent reads compact text, reduced on your computer before anything is sent.
Tokenade helps people who work their documents with an AI agent running on their own computer: Claude Code (including the Code tab of Claude Desktop), Codex, Cursor and the other agents in our compatibility list. It turns the files and pages the agent reads into compact text, on your computer, before they reach the model.
Why these roles use AI agents
Analysts, lawyers, consultants and marketers do not write software, but their work has the same shape as a developer's: a folder of inputs, a set of repetitive steps, and a structured output. An agent that can open every file in a folder, follow a checklist and write the result is a natural fit. That is why Claude Code, first built for developers, now shows up in contract review, financial reporting, client decks and SEO audits.
Where the tokens go
For these roles, the bill is not driven by code. It is driven by reading:
- PDFs. Anthropic's documentation counts 1,500 to 3,000 tokens per PDF page for the text, with an image of each page billed on top (Claude PDF support, read 2026-10-15).
- Office files. Word, Excel and PowerPoint files are bundles of XML and layout that an agent can end up reading along with the content.
- Web pages. Raw HTML carries navigation and scripts that say nothing useful.
- Connectors. Analytics, contract or market-data tools plugged in over MCP can return long tables.
Every file the agent has read stays in the conversation and is sent again on each turn. And on Claude plans, Claude and Claude Code draw from the same usage allowance (Claude for Lawyers, 2026-09-09, citing Anthropic's Help Center), so a heavy session in one leaves less for the other.
Tokenade replaces that raw material with extracted text, lets the agent ask a long document questions instead of reading it whole (documents and media), and folds long tool results. On our public benchmark it makes long agent sessions 39% cheaper.
Which Claude apps it works with
Tokenade works where the agent runs on your computer, because that is where it can step in between the agent and your files.
- Claude Code and the Code tab of Claude Desktop: fully supported.
- Claude Desktop chat: only results from MCP connectors are optimised. The chat does not open your files through local tools, so there is nothing else for Tokenade to compact.
- Claude Cowork: not supported. Cowork runs in an isolated environment that local tools cannot reach.
If you work in Cowork today and want Tokenade's savings, the Code tab of Claude Desktop is the closest path: it works on folders on your own computer. Codex and Cursor are covered too.
What these roles have in common
The files are confidential, and they are long. Tokenade's compaction runs on your computer, and your documents do not pass through Tokenade's servers; only usage counts are sent to meter your plan. Tokenade masks credentials such as API keys, but it does not remove names or client details from your documents. The pages below go into what that means for each role.