Claude Code for research: more papers per session, fewer limit hits

You put forty papers in a folder, ask your AI agent for an extraction table, and the usage bar is gone before the synthesis starts. PDFs are the most expensive thing a research workflow feeds an agent, and a literature review feeds it dozens. Tokenade turns those papers into compact text on your computer, so more of them fit in a session and in your plan.

Figures on this page: Anthropic's PDF documentation and practitioners' guides, read on 2026-10-15; Tokenade's own figure comes from the published THOL benchmark.

What a literature review does to your agent

Research use of Claude Code is well documented now: screening abstracts, building extraction tables, drafting themes from a folder of sources. A systematic-review workflow with dedicated /screen and /extract commands is published by research-memex (2026-05-14). The friction shows up in four places.

Every page is billed as text and as an image

Anthropic's PDF support documentation (read 2026-10-15) states that each page "typically uses 1,500–3,000 tokens per page depending on content density", and that because "each page is converted into an image", image costs apply on top. A 30-page paper is a large share of a session before the agent has written a line.

A bigger context window does not fix it

When Claude Code moved to a 1M context, Aniket Panjwani tried a literature review of as many Daron Acemoglu papers as would fit. His finding (X, 2026-03-14): "CC will try to "get around" actually fitting the full text of papers into its context". His write-up on ai-mba.io concludes that "PDFs still trigger hidden bottlenecks, auto-compaction still matters, and preprocessing still wins when coverage matters."

Long papers are read in slices

The CC for Everyone research guide (updated 2026-06-10) notes that the agent reads anything over 10 pages "in ranges, up to 20 pages per request", and that scanned PDFs need OCR first. A 60-page working paper means several reads, each one added to the conversation.

The context fills and quality drops quietly

The same guide warns against loading a whole corpus at once: "The context window fills, quality sags, and you cannot tell where." For a review whose value is coverage, silent degradation is the worst kind of failure.

On top of this, Claude's limits are shared: the 5-hour window and the weekly cap are drawn by chat, Claude Code and Cowork alike (claudelimit.com, 2026-09-01). Tokens spent on page images are tokens you cannot spend on writing.

What Tokenade changes for a literature review

Tokenade sits between your AI agent and its tools on your computer. For research work, five features do most of the job.

• Papers arrive as text. When the agent opens a PDF, DOCX or EPUB, Tokenade's read hook returns the extracted text instead of page images. Scanned pages are OCR'd when tesseract and pdftoppm are installed. See Read documents and media.
• Ask a paper instead of reading it. tokenade read paper.pdf --prompt "sample size, identification strategy, main effect" returns only the passages that answer, verbatim, each under its own heading. Your agent learns this command from the rules tokenade install writes.
• No paying twice for the same paper. When the agent re-opens a file that has not changed, it gets a short pointer instead of the full text again.
• Nothing is thrown away. Long outputs are folded, and the complete text stays in a local stash the agent can query or restore. Exact quotes remain checkable.
• Lighter web searches. Journal pages, arXiv listings and search results are compacted before they reach the model.

Across long agent sessions, Tokenade is 39% cheaper than running without it, measured on the THOL benchmark. On a Pro or Max plan, that shows up as more papers per window rather than money back. tokenade gain shows what was saved on your own work.

What stays on your computer

Extraction, OCR and compaction run on the computer where you work. No paper, draft or prompt passes through Tokenade's servers. What Tokenade sends is usage counts: token numbers, the compactor's name, the agent and the model, never file names or contents (full list).

Be precise about the rest. Your agent still sends what it reads to its own provider, Anthropic or OpenAI for example, so a confidential manuscript is shared with that provider either way. Tokenade masks credentials it recognises, such as API keys in a script. It does not hide authors' names or reviewers' comments.

Tokenade works with agents that run on your computer: Claude Code (including the Code tab of Claude Desktop), Codex, Cursor and others. It has no effect in Claude Cowork, and in the Claude Desktop chat only MCP results are optimised (compatibility).

Setting it up

1. Install the CLI and run tokenade install. It finds Claude Code, Codex, Cursor and the other supported agents on your computer and wires each one (install guide).
2. Optional: install tesseract and pdftoppm for scanned PDFs, so old chapters and photocopies come back as text too.
3. Open your review folder in your agent as usual. Ask for the extraction table, then run tokenade gain to see how many tokens the papers no longer cost.

FAQ

Does Claude Code not read PDFs already?

It does, by sending each page as an image plus its text. Tokenade's read hook replaces that with the extracted text, and lets the agent ask a long paper for specific passages instead of loading it whole.

Will I lose exact quotes or table values?

No. --prompt returns passages verbatim, and anything folded stays in the local stash, where the agent can restore the full text. Complex tables can still come out garbled from any text extraction, so check critical numbers against the page.

Does it work in Claude Cowork or the Claude app chat?

Not in Cowork, which runs in an isolated environment local tools cannot reach. In Claude Desktop chat, only MCP results are optimised. Use Claude Code, including the Code tab of Claude Desktop, or another supported agent.

Is my unpublished manuscript sent to Tokenade?

No. Compaction runs locally and Tokenade only receives usage counts. Your AI agent's provider still receives whatever the agent reads, as it would without Tokenade.

Does it help with citation hallucinations?

Not directly. It keeps the full text of each source reachable, which makes the verification pass cheaper to run on every quote. The check itself is still yours.

Further reading

• Claude usage limits: how they work
• How the context window drives your token bill
• How to reduce tokens on long agent sessions
• The THOL benchmark and its methodology