Docusense
本地 MCP 服务器和 Claude Code 插件,可为 PDF、DOCX、HTML 和 Markdown 文件建立索引,让 Claude 按章节和论点读取内容,而不是加载整个文件。每段内容都附带经过 SHA-256 验证的引用;重新添加的文件会成为新版本,并支持差异对比、责任追溯和跨文档矛盾检查。
文档
DocuSense
Stop pasting whole documents into your agent.
DocuSense is a local MCP server that lets Claude, Cursor and other AI agents read your PDFs, Word files and docs by section and by claim, so they load the few hundred tokens that answer the question instead of the whole 50,000-token file. Every answer comes with a citation to the exact passage.
pasting the docs ████████████████████████████████████████ 160,179 tokens
vector RAG ████ 15,765 tokens
DocuSense ██ 6,869 tokens → 23x less than pasting
Install
uv tool install docusense-mcp # or: pipx install docusense-mcp
docusense setup # connects to Claude Code
docusense add ~/docs contract.pdf # files or whole folders
That's it. Ask your agent about your documents.
Other clients: docusense setup claude-desktop, docusense setup cursor, or docusense setup print
for a config snippet to paste anywhere. Claude Desktop users can also install the one-click
docusense.mcpb from the releases page.
Supported files: PDF, Word (.docx), HTML, Markdown, plain text.
See what you saved
Example:
$ docusense savings
12 agent sessions · 87 tool calls
read via DocuSense: 41,302 tokens
pasting the docs: 912,440 tokens
saved: 871,138 tokens (22.1x less context)
Why it uses fewer tokens
| You ask | Pasting the docs | Vector RAG | DocuSense |
|---|---|---|---|
| A precise fact, with a citation | 9,090 tokens | 5,362 tokens of nearby chunks | 478 tokens: the exact paragraph |
| What changed between two versions | 97,895 tokens | 5,292 tokens, can't diff | 2,776 tokens: just the changes |
| Does document A conflict with B? | 53,194 tokens | 5,111 tokens, hit or miss | 3,615 tokens: both conflicting passages |
- Reads like a person. Your agent sees the table of contents, opens the right section, and reads only the paragraph it needs.
- Tracks versions. Add a changed file again and DocuSense keeps both versions. Ask what changed without loading either one.
- Finds contradictions between documents, with both sides cited.
- Citations you can trust. Every passage is fingerprinted, so a made-up quote can't pass as a real one.
- Private and free. Runs on your machine. No API key, no account, nothing uploaded.
Great for
- Long specs, API references and design docs while coding in Claude Code or Cursor
- Checking policies against standards and tracking revisions
- Contracts, handbooks and research papers where you need exact quotes and page numbers
- Any agent that keeps re-reading the same large files
If your documents are a few short files that already fit in context, you don't need DocuSense.
What your agent can do
| Browse | list documents, view the table of contents, read a section |
| Find | search, find requirements and rules, look up definitions, follow cross-references |
| Cite | get a verified quote, pack evidence into a token budget |
| Compare | list versions, diff versions, see when a line changed, find contradictions, find affected documents |
| Add | add files on its own and track progress |
Set DOCUSENSE_TOOLS=core for a smaller tool set if your agent's context is tight.
Settings
| Variable | Default | |
|---|---|---|
DOCUSENSE_HOME | ~/.docusense | where your data lives |
DOCUSENSE_ADD_ROOTS | your home folder | folders the agent may add files from |
DOCUSENSE_TOOLS | all | core for fewer tools |
DOCUSENSE_SAVINGS_LOG | 1 | 0 turns off the local savings log |
Share one library across several agents: docusense serve-mcp --http (set DOCUSENSE_HTTP_TOKEN
when exposing it beyond localhost).
The first time you add documents, DocuSense downloads a small local search model (~130 MB) once.
License
MIT