DocSift

Convert documents once. Give agents only what they need.

mcp-name: io.github.anishmoncivarghese/docsift

A 300-page PDF does not fit in a language model's context window, and pasting it in would be expensive if it did. DocSift converts documents into clean Markdown once, indexes them, and then hands back only the passages that answer a question — with page numbers and section headings, so the answer can be cited.

It runs on your machine. PDFs go through Docling, everything else through MarkItDown, both behind one interface. No cloud APIs, no accounts, no telemetry.

Where the saving comes from: retrieval, not conversion. Cleaning barely reduces tokens on PDFs, because Docling already strips headers and footers with its layout model. What changes the bill is asking a question and getting three relevant chunks back instead of a whole document.

Quickstart

Pick the engine you need — markitdown for Word, Excel, PowerPoint, HTML, CSV and EPUB; docling for PDFs (a large download: ML layout models); all for both engines plus the HTTP API and MCP server.

pip install "docsift[markitdown]"
pip install "docsift[docling]"
pip install "docsift[all]"

docsift convert report.pdf

That writes cleaned Markdown, token-budgeted chunks and a JSON summary to ./output/. pip install docsift on its own installs the CLI but no engine, and conversion will tell you so rather than failing obscurely.

Office means the modern formats: .docx, .xlsx, .pptx. The pre-2007 binaries — .doc, .xls, .ppt — are not supported, and DocSift refuses them by name rather than failing halfway through. Open one in PowerPoint, Keynote, LibreOffice or Google Slides and save it as .pptx first.

Use it from Claude, Codex, or another MCP client

The shortest path to the point of this tool: let an assistant search your own documents, without pasting them anywhere.

Which clients this works with

DocSift speaks MCP over stdio — the client starts it as a program on your machine. Anything that can do that is supported:

The last two are worth being clear about before you install anything. Their connector features only accept a remote MCP server at a public HTTPS URL, and DocSift has no remote transport — there is no configuration that makes a local one appear in those interfaces.

That is a deliberate position rather than an oversight. Reaching them means running DocSift on a server and uploading your documents to it, which is the opposite of the thing this tool is for. A self-hosted remote transport is a reasonable future addition; sending your files to someone else's machine is not.

1. Install it as a command, not into a project

An MCP client starts DocSift as a program, so it has to exist outside any virtualenv. Install it as a standalone tool:

uv tool install --python 3.12 "docsift[mcp,docling,markitdown]"

or with pipx:

pipx install --python python3.12 "docsift[mcp,docling,markitdown]"

DocSift needs Python 3.11 or newer. If your default is older, the version flag above is what avoids an unsatisfiable-requirements error. Expect a large download: docling brings PyTorch and layout models.

On Linux, add --torch-backend auto. The default resolves to the CUDA build of PyTorch — 5.3 GB installed, roughly 2 GB of it nvidia-* wheels that a machine without an NVIDIA GPU never loads. auto detects your driver and picks the right build, which is 1.6 GB on a machine without a GPU and leaves CUDA in place on one with a GPU:

uv tool install --python 3.12 --torch-backend auto "docsift[mcp,docling,markitdown]"

macOS wheels are CPU-only already, so the flag changes nothing there. It is a uv feature: with pipx or pip there is no equivalent, because the CPU builds live on a separate index and no published package can redirect an installer to it. If you install that way on a CPU-only Linux box, DocSift says so after your first conversion rather than letting several unused gigabytes pass unmentioned.

Check it landed. Run these one at a time; the second prints the path to the executable, which the next step needs.

docsift --version

which docsift

2. Register it with your client

Pick the one you use. You only need one of these.

Claude Code

$(which docsift) fills in the path for you, so this works exactly as written:

claude mcp add --scope user docsift -- "$(which docsift)" mcp

Then confirm it started — look for docsift ... ✔ Connected:

claude mcp list

--scope user makes it available in every project; without it, the server is registered only for the directory you were in.

That is the whole setup for Claude Code. Skip the other clients below and go to step 3.

VS Code

Copilot agent mode reads .vscode/mcp.json for one project, or the file behind the MCP: Open User Configuration command for all of them:

{
  "servers": {
    "docsift": {
      "type": "stdio",
      "command": "/replace/with/the/path/from/which/docsift",
      "args": ["mcp"]
    }
  }
}

Codex CLI

~/.codex/config.toml:

[mcp_servers.docsift]
command = "/replace/with/the/path/from/which/docsift"
args = ["mcp"]
startup_timeout_sec = 60

Raise the timeout as shown. Codex allows ten seconds by default and DocSift loads PyTorch on the way up, so the default reports a server that failed to start when it was only still starting.

Claude Desktop, Cursor

These read a JSON config file instead. For Claude Desktop on macOS that is ~/Library/Application Support/Claude/claude_desktop_config.json; create it if it does not exist.

If the file is empty or new, this is the whole contents — replacing the command with the path which docsift printed in step 1:

{
  "mcpServers": {
    "docsift": {
      "command": "/replace/with/the/path/from/which/docsift",
      "args": ["mcp"]
    }
  }
}

If it already has other servers, add docsift beside them rather than replacing the file — note the comma after the previous entry:

{
  "mcpServers": {
    "something-you-already-had": {
      "command": "..."
    },
    "docsift": {
      "command": "/replace/with/the/path/from/which/docsift",
      "args": ["mcp"]
    }
  }
}

Use the absolute path, not a bare docsift. These clients do not reliably inherit your shell's PATH, and a wrong or bare path fails with ENOENT: no such file or directory.

Then restart the app — the config is read at startup.

Use the absolute path from which docsift, not a bare docsift. MCP clients do not reliably inherit your shell's PATH, and this is the most common reason a local server silently fails to start.

3. Ask

No commands to learn — describe what you want:

search ~/Documents/contract.pdf for the termination clause

what does report.pdf say about Q3 revenue?

The first question about a new file converts it, and on a PDF that is slow — about three minutes. That is startup cost, not page count: Docling downloads its layout and table models from HuggingFace on the very first conversion, then loads PyTorch. A three-page test file takes about as long as a thirty-page report, so picking something small to "try it quickly" does not help.

It happens once. Afterwards the file is recognised by its content and answers come back immediately, even if you move or rename it.

For a long PDF, convert it first and ask afterwards:

docsift convert big-report.pdf

That fills the same cache the MCP server reads, so the first question is as fast as the rest. It also sidesteps a real limit — MCP clients apply their own timeouts to a tool call, and a long enough conversion can exceed one and surface as an error even though it would have finished.

4. Check it works

Before pointing it at a real PDF, prove the wiring with a file that converts instantly. Make one:

printf '# Test\n\nDocSift returns only the passages that match a question.\n' > /tmp/docsift-check.md

Then ask your assistant, naming the tool:

use the docsift search_document tool on /tmp/docsift-check.md to search for "passages"

Name it explicitly, and watch which tool actually runs. Asked casually, an assistant will often just open a small file with its own file-reading tool and answer from that — you get the right answer having never touched DocSift, which makes a broken setup look like a working one. If your assistant reports reading a file rather than calling docsift, the check has told you nothing.

The first call will ask your permission to run the tool — approve it, and choose the "don't ask again" option if your client offers one, so later questions are not interrupted mid-thought.

Success looks like a docsift tool call in the transcript, returning that one sentence. This needs no PDF and no model weights, so a failure here is a setup problem — the command not on PATH, or the server not registered — and not a conversion one.

Then try a real PDF of your own, and expect the first question to take a minute. On a document of that size the choice takes care of itself: reading it whole is expensive, which is when searching it becomes the obvious move.

One test worth running deliberately: ask about something the document does not mention. You should get nothing back rather than a plausible-sounding answer. Search is lexical, so a word that is not in the text matches nothing — and seeing that once tells you more about what you can trust than a dozen successful queries.

What you get

Two tools:

• search_document — the one that matters. Give it a file path and a question; it converts the file the first time it sees it, then returns only the passages that match, with page numbers and section headings.
• convert_document — converts and indexes a file, returning a summary (page count, token estimate, chunk count) rather than the text.

What to expect

Measured on an Apple-silicon MacBook Air with a 34-page, 1.7 MB PDF — a BusinessEurope economic outlook, chosen because it is the awkward kind of document this tool exists for: dense tables, footnotes, multi-column stretches.

The token gap is the point, and it widens with document length.

Be clear about where the three minutes goes, because it is easy to misread as "big documents are slow". Most of it is Docling fetching its layout and table models the first time it ever runs, plus loading PyTorch. On a clean Linux machine a three-page, 1.8 KB PDF took 186 seconds — essentially the same as the 34-page report. After that first run the models are on disk and conversion scales with the document. docsift convert shows a live spinner throughout, so you can see it working rather than guessing.

search_document takes limit and max_tokens as well, and a model will set them when you ask for more or less. The defaults return five passages within a 5,000-token budget. Dense documents produce large chunks — around 1,000 tokens each in the report above — so if answers feel truncated, a larger max_tokens is the dial, and it is cheaper than the model asking the same question several times over.

One caveat the numbers do not show: search is lexical, not semantic. DocSift indexes chunks in SQLite FTS5 and ranks them with BM25. It matches the words you type — not synonyms, not paraphrases, not meaning. Ask for "termination clause" and a section that only ever says "ending the agreement" will not come back. Asking that report about "energy prices" surfaced a section on shipping and fertiliser costs because those passages happen to contain the phrase; a question worded "why did transport costs rise?" might have missed it entirely.

That is a deliberate trade, not an oversight: no embedding model to download, no index to rebuild, no GPU, nothing leaving your machine, and a first run measured in minutes rather than tens of minutes. If you need semantic retrieval, the Markdown and chunk JSON DocSift writes are clean input for a vector store — a reasonable thing to want, and not what this tool does today. See Known limitations.

PowerPoint answers cite slide numbers. A deck is chunked one slide at a time, so a passage comes back saying which slide it is from and what that slide was titled — the same way a PDF passage carries its page. The field is called pages for both; for a presentation it means slides. Speaker notes are extracted too, and are searchable.

What a deck will not give you: diagrams. A slide with three boxes and two arrows extracts as three labels. The words survive, the relationships do not, so DocSift cannot tell you what an arrow between two boxes meant. Images inside slides are not read at all — text in a screenshot is invisible to search.

Everything happens in that process, on your machine. Nothing listens on a port and no document content crosses the network. The server has exactly the filesystem access of the user who started it.

Not in the claude.ai connector directory, and it cannot be. claude.ai runs in the cloud and cannot start a program on your computer. A local MCP server works in apps running on your own machine — Claude Code, Claude Desktop, Codex, Cursor — and searching the web app's connector list for DocSift will never find it.

Command line

docsift convert report.pdf --max-tokens 800 --overlap 100
docsift convert report.pdf --engine markitdown
docsift inspect report.pdf
docsift compare report.pdf
docsift search doc_xxxxxxxxxxxx "operational risk"
docsift search doc_xxxxxxxxxxxx '"operational risk"' --limit 5 --context 1
docsift cache info
docsift cache clear

inspect reports what DocSift would do with a file — engine, identity, cache status — without converting it. compare runs both engines on the same document and writes a diff of the results.

Conversion cleans repeated headers and footers, page numbers and image references, then splits the text into token-budgeted chunks that carry their heading context. --keep-furniture and --keep-image-refs turn the cleaning stages off.

Results are cached, so an unchanged file with unchanged settings returns instantly. --no-cache forces a re-run.

Search is local SQLite FTS5 keyword ranking with quoted-phrase support. It reads the store the HTTP service and the MCP server write to — docsift convert writes standalone files to an output directory and does not add them to it. --context pulls in adjacent chunks, --max-tokens caps the whole response, and only selected chunks are ever printed. Scores order results within one response and are not comparable across requests.

HTTP API

pip install "docsift[all]"
docsift serve

Conversion always runs in the background: a long PDF can take minutes, and a client expecting a synchronous response will time out.

Upload, and get back 202 with a job id and a document id:

curl -sS -F file=@report.pdf http://127.0.0.1:8000/v1/documents

Poll until the status is succeeded or failed:

curl -sS http://127.0.0.1:8000/v1/jobs/job_xxxxxxxxxxxxxxxx

Then retrieve the whole document:

curl -sS http://127.0.0.1:8000/v1/documents/doc_xxxxxxxxxxxx/markdown
curl -sS http://127.0.0.1:8000/v1/documents/doc_xxxxxxxxxxxx/chunks

Or ask a question and get only the relevant chunks:

curl -sS --get --data-urlencode 'q=operational risk' \
  --data 'limit=5' --data 'max_tokens=5000' \
  http://127.0.0.1:8000/v1/documents/doc_xxxxxxxxxxxx/search

DELETE /v1/documents/{id} removes the document, its search index and its cached conversions — including cancelling a conversion still in flight, so a delete cannot be undone by a worker finishing afterwards. The OpenAPI document is at /openapi.json.

Set DOCSIFT_API_KEY before anything else can reach it. Every /v1/* route then requires an X-API-Key header, while /health, /version and the API docs stay open for health checks and connector imports. It is one shared secret for the whole service — not per-user identity, and no substitute for network controls. The service converts whatever it is given, so run it on infrastructure you control.

Power Platform, Copilot Studio and n8n

DOCSIFT_PUBLIC_URL=https://docsift.example docsift openapi --format swagger2 -o connector.json

Import that as a Power Platform custom connector. The service's own /openapi.json is OpenAPI 3.1, which custom connectors reject; this command emits the Swagger 2.0 they accept. Verified against a real tenant: the file imports, authenticates and returns schema-valid responses.

Worked examples live in examples/ — an importable n8n workflow, a Copilot Studio connector walkthrough, and the Power Automate Do until flow that waits for a conversion. A longer guide is in docs/USING_DOCSIFT.md.

More

License

MIT