MCP Server
17 tools your agent calls to browse, search, cite and compare documents. Runs on stdio or HTTP.
Agents that quote your documents,
at a fraction of the tokens
A local MCP server that reads your PDFs, Word files and specs by section and by claim. Every answer cites the exact passage.
$ uv tool install docusense-mcp
Yes. v2 says a line MUST NOT exceed 1000 octets, matching RFC 5321 §4.5.3.1.6.
✓ ds://mail-policy@v2#§3.1:b812-951
✓ ds://mail-policy@v2#§3.1:b812-951
@mail-policy@rfc5321
Tools →
CLI
Add files or whole folders, search from the terminal, and run docusense savings to see what you saved.
Ledger
An append-only store. Every passage has a byte range and a SHA-256, and old versions are never overwritten.
23x
Less context than pasting
478
Tokens for a cited fact
100%
Passages hash-verified
0
API keys or accounts
Reads the files you already have
PDF
Word
HTML
Markdown
Text
#01 - Read like a person
Table of contents, then the one paragraph that answers.
› get_section_map(doc="mail-policy")
mail-policy@v2 · 6 sections · 1,412 tokens
§1 Scope96
§2 Definitions188
§3 Message Limits402
§3.1 Line Length121
§3.2 Recipients104
§3.3 Local-part Length98
§4 Delivery Retries231
§5 Command Support142
§6 Transport Security117
outline only · body not loaded
#02 - Citations you can check
Every quote is fingerprinted, so a made-up one can't pass.
› get_evidence(span_id=48)
A message line MUST NOT exceed 1000 octets, including the CRLF.
stored4be1c9d2…a3e107af
recomputed4be1c9d2…a3e107af
versionv2 · current
page2
#03 - Versions & conflicts
Ask what changed. Ask what contradicts. RAG can't.
› diff_versions(doc="mail-policy")
v1 → v2 · 4 changed · 1 added
§3.1 Line Length
− A message line MAY be up to 2000 octets+ A message line MUST NOT exceed 1000 octets§3.2 Recipients
− MUST NOT accept more than 50 recipients+ MUST accept at least 100 recipients§6 Transport Security
+ A relay server MUST offer STARTTLS#04 - How it works
Four commands from a folder of files to cited answers.
01
Install
One Python package. No server, no account, no API key.
$ uv tool install docusense-mcp
02
Connect
Registers the MCP server with Claude Code. Cursor and Claude Desktop work too.
$ docusense setup
03
Add
Files or whole folders. Re-add a changed file and you get a new version.
$ docusense add ~/docs
04
Ask
Ask your agent about your documents. Then check what it saved you.
$ docusense savings
Under the hood
Parsed once, by rules. Indexed locally. Never overwritten.
Parsing and extraction are deterministic. The search model (~130 MB) downloads once and runs on your machine. Nothing leaves it.
Quickstart →Parsepdf · docx · html · md
Clean text, page map, and a section tree
Spansbyte range + sha256
Paragraphs, tables and sentences, each fingerprinted
Extractrules first
Claims like
MUST MUST NOT, definitions, referencesIndexbm25 + embeddings
Fused ranking, served to your agent over MCP
Benchmarks
Same questions. A fraction of the tokens.
less context than pasting the docs
23x
A precise fact
with a citation
What changed
between two versions
Does A conflict with B?
both sides cited
Tokens of context the agent loaded per question, measured with scripts/token_demo.py. Totals: 160,179 pasted · 15,765 vector RAG · 6,869 DocuSense. Measure your own →
Agentic integrations
Works with your favorite agent.
Parsing and extraction are rules, not guesses. An LLM fallback exists, but it's off by default and every call it makes is recorded.
Deterministic firstLLM optional
Old evidence is never overwritten. The database itself rejects edits and deletes to versions, sections and spans.
Append-onlyEvery version kept
Embedded SQLite and a local search model. No API key, no account, nothing uploaded. If your files already fit in context, you don't need it.
Local & privateNo account needed
Still wondering?
Ask your favorite LLM what it thinks of DocuSense, then make an informed decision.