MCP server for Claude

Your research folder, typeset into a paper.

PaperPilot reads your code, data, and notes, resolves the venue's real formatting rules, drafts and compiles a LaTeX paper, then checks it for hallucinated claims. Runs entirely on your machine — nothing is ever uploaded.

Download PaperPilot.exe See how it connects

research/

code/

early_exit.py

baselines.py

calibration.py

data/

main_results.csv

ablation_threshold.csv

dataset_stats.csv

figures/

accuracy_vs_k.png

latency_scatter.png

notes.md

> screening done

> venue_resolution done

> structure_drafting done

> humanization

→

main.tex · compiled

Confidence-Gated Early Exit for Efficient Classifier Inference

Anonymous Author(s)

Affiliation withheld for double-blind review

Abstract—Standard deep classifiers run every layer on every input, even when the network is already confident after a few layers. We propose confidence-gated early exit, which halts inference once per-layer confidence clears a threshold, requiring no architectural changes or retraining.

1 Introduction

A deep classifier with Llayers commits the same compute to every input, whether it is easy or hard to classify. Consider an input the network could confidently classify after four layers—it still pays for all twelve, because the standard forward pass carries no mechanism for stopping early [1].

We study a simple mechanism for addressing this waste. At each layer of an already-trained classifier, we compute the softmax confidence of the current layer's prediction. Once this confidence exceeds a fixed threshold τ, inference stops immediately and the current prediction is returned [2].

Across three benchmark datasets, confidence-gated exit retains 98.6–99.3% of full-network accuracy while running only 61–76% of layers on average, closing most of the gap to an oracle upper bound.

Why it's built this way

Claude doesn't touch your files. PaperPilot does.

Claude Desktop and claude.ai run in the cloud — they have no way to reach a folder on your laptop. So instead of uploading your research anywhere, PaperPilot runs locally and opens a private tunnel just so Claude can ask it to do things. Every read, every LaTeX compile, every file write happens on your machine, under a token only you have.

Claude Codeskips the tunnel entirely — it already runs locally, so it talks to PaperPilot directly, no public URL involved at all.

Claude

→private tunnel

→PaperPilot (your PC)

→your files · Tectonic

Claude only ever sees what PaperPilot's tools return to it — never a direct line to your disk.

One download, three ways in

Connect it however you already use Claude

Claude Desktop

Double-click PaperPilot.exe. It opens a private tunnel and prints a URL — paste it into Settings → Connectors.

Remote connector

claude.ai (web)

Identical to Desktop — same exe, same tunnel, same URL. Add it as a custom connector in your browser.

Remote connector

Claude Code

Runs on your machine already, so no tunnel is created. Register the exe directly with --stdio.

Local, direct

Try it now:

1. Surface→2. Commands→3. Verify

Which do you use?

Ready to try it?

One executable, no account, and nothing leaves your machine.

Download PaperPilot.exe