Consultants: decks, data packs and desk research that burn through your AI agent's tokens

A consulting engagement starts with a pile of inputs: the client's data pack, last year's decks, interview notes, a dozen industry reports and as many web pages. Hand that to an AI agent and most of the tokens go on reading, before any synthesis is written. Tokenade turns those inputs into compact text on your computer before the agent sees them, so the reading costs less and the session lasts longer.

This page is for consultants who already work with an AI agent that runs on their computer: Claude Code (including the Code tab of Claude Desktop), Codex, Cursor and similar. Sources below were read on 2026-10-15.

What you deal with every day

Consultants who use Claude Code describe the same pattern: many inputs in, one structured deliverable out. One independent consultant built a script that "takes raw research inputs (competitor websites, client interview notes, analytics exports) and generates a formatted report" (Alejandro Arce, 2026-06-19). Every one of those inputs is paid for in tokens.

Decks and Word files carry weight you cannot see

A .pptx or .docx file is a zipped bundle of XML, styles and layout. As one widely shared guide on Claude limits puts it: "DOCX and PPTX files carry metadata bloat you can't even see" (Ruben Hassid, 2026-04-12). A slide-tools vendor makes the same point about PowerPoint: "PowerPoint files are XML-heavy, and Claude must process all of it" (auxi, 2026-05-20).

Reports and data packs arrive as PDFs

Anthropic's documentation says a PDF page "typically uses 1,500–3,000 tokens per page depending on content density", with an image of each page billed on top (Claude PDF support, read 2026-10-15). Three analyst reports of sixty pages each can fill much of a session on their own.

Desk research means raw web pages

A web page in HTML carries navigation, scripts and markup that say nothing about the market you are studying. Cloudflare measured its own announcement post at 16,180 tokens in HTML against 3,150 tokens as markdown (Cloudflare, 2026-02-12). Twenty competitor pages read raw add up fast.

Client confidentiality is part of the brief

Firms take NDAs seriously when AI enters the workflow. A training firm's guide for consulting teams recommends a Team or Enterprise plan with a signed data processing agreement, and removing identifying data before anything is submitted (formation-claude-ia.fr, 2026-05-29). Any tool added to the chain has to fit that policy.

What Tokenade changes for consulting work

Tokenade sits between your AI agent and the files, pages and tools it uses. It changes what the model reads, not what it writes: your deck is still built the way you build it today.

• Decks, documents and workbooks become text. When the agent opens a .pptx, .docx, .pdf or .xlsx, Tokenade returns the extracted text instead of the file's internals and page images.
• Ask the report, don't read it. tokenade read market-study.pdf --prompt "market size, growth drivers, main competitors" returns only the passages that answer, each under its own heading, in one call (how it works).
• Interview recordings become transcripts. With whisper installed, an interview recording is transcribed on your computer and read as text.
• Web pages come back as compact markdown. tokenade web fetches pages without the markup, and tokenade search runs one query across several search engines and returns titles, links and snippets (web and search).
• No paying twice. When the agent re-reads a file it already saw, Tokenade spots the repeat.
• Savings you can show. tokenade gain lists what was saved, row by row, which helps when the firm asks what the AI budget buys.

On our public benchmark, Tokenade makes long agent sessions 39% cheaper. A week of synthesis on one engagement is a long session.

What stays on your computer

Compaction runs on the computer where you work. Tokenade has no telemetry: client files, prompts, file names and usage figures stay on your machine, and Tokenade sends nothing anywhere (What Tokenade sends and stores).

Two points to be clear about with your firm. The compacted text still goes to your AI provider, as it would without Tokenade, so your engagement terms on that provider still apply. And Tokenade masks credentials it recognises, such as API keys and tokens, but it does not remove client names, project names or figures. If your NDA requires that, it has to happen before the files reach the agent.

Setting it up

You need an AI agent that runs on your computer. Claude Code, the Code tab of Claude Desktop, Codex and Cursor all work. Claude Cowork is not supported, and in the Claude Desktop chat only MCP results are optimised (compatibility).

1. Install the command line tool: npm install -g @tokenade/cli.
2. Run tokenade install. It finds your agents and connects to each one.
3. Work on an engagement folder as usual, then check tokenade gain.

Full instructions are in Install Tokenade. Tokenade is free, with no account to create; for a whole practice, each consultant runs the same two commands.

FAQ

Does Tokenade change the decks my agent produces?

No. Tokenade compacts what the agent reads: files, web pages and tool results. The slides, documents and spreadsheets the agent writes are created exactly as before.

Does it work in Claude Cowork or Claude's slide tools?

Not in Cowork, which runs in an isolated environment that local tools cannot reach. Tokenade works when an agent on your computer, such as Claude Code or Codex, opens your files. What it saves there leaves more of your shared Claude allowance for the rest.

Do client files leave my computer?

Not to Tokenade. Compaction is local and Tokenade sends nothing. The text your agent works with goes to your AI provider, as it would without Tokenade.

Can my whole firm use it?

Yes. Each consultant installs it on the computer where they work. There is no account, licence or seat to manage, and nothing passes through a Tokenade server.

For other roles that work mostly with documents, see the document work overview.