Skip to main content
Glama
PaperPilot

PaperPilot

by sjgod1427

PaperPilot

An MCP server that turns a research folder — code, data, figures, notes — into a compiled, submission-ready academic paper. Runs entirely on your machine; nothing is ever uploaded.

Website: sjgod1427.github.io/PaperPilot Download: Latest release

What it does

Point PaperPilot at a research folder and it runs a five-stage pipeline:

  1. Screening — checks the research for issues that would sink a submission (test-set leakage, statistical-validity problems, illegible figures, cross-file numeric mismatches) before writing a word.

  2. Venue resolution — figures out the target venue's real formatting rules and page limits, or helps you pick one.

  3. Structure & drafting — writes the paper into a real LaTeX template, filling every page it's allowed to.

  4. Humanization — rewrites the draft for plain, direct academic prose.

  5. Final QA — checks the compiled paper against the source material for hallucinated claims before calling it done.

Related MCP server: ares-mcp

Using it

Download PaperPilot.exe from the latest release and run it — see INSTALL.md for the full setup for Claude Desktop, claude.ai, or Claude Code. The website walks through the same steps interactively, including a live connection checker.

Repository layout

  • src/paper_writing_pipeline/ — the MCP server (Python, uv-managed).

  • website/ — the marketing/distribution site (Next.js, static export, deployed to GitHub Pages via .github/workflows/deploy-website.yml).

  • tests/ — Python test suite (uv run pytest).

Development

Python side:

uv sync
uv run pytest -q

Website:

cd website
npm install
npm run dev

Available Tools

21 tools
add_template_to_libraryA

Add a new template to the user's library (never writes into the bundled seed dir).

Always sets metadata["id"] to template_id, overriding whatever (if
anything) the caller passed for "id" -- this is what get_template_files
matches on, so a missing or mismatched "id" would silently make the
template unreachable (or crash list_templates() entirely if "id" is
missing), which happened for real once already.

Inputs: files (required), metadata (required), template_id (required), user_library_dir.

append_project_fileB

Append to a plain, user-visible file at the project root, creating it if missing.

Inputs: content (required), filename (required), project_dir (required). Returns: result

cleanup_rendered_pagesB

Remove the rendered page-image folder once it's no longer needed.

Inputs: pdf_path (required). Returns: result

compile_latexA

Compile main.tex in project_dir with Tectonic.

pdf_path is None if compilation failed to produce a PDF. log_text always
contains the full compiler output so the caller can diagnose and fix the
source before retrying.

Inputs: project_dir (required).

copy_fileA

Copy a file byte-for-byte -- for binary files (figures, PDFs) that read_file/write_file can't safely round-trip since they decode as UTF-8 text. Use this to bring a figure from research_folder into project_dir so the paper references it with a relative path, not an absolute path pointing outside the project -- a real gap found in practice: without this, a paper compiled fine but only by \includegraphics-ing a figure from its original external location, which isn't a self-contained submission package.

Inputs: dest_path (required), source_path (required). Returns: result

final_qa_promptC

Stage 4 instructions: whole-paper consistency/hallucination/venue-fit check.

Inputs: venue_name (required), project_dir (required), max_roundtrips. Returns: result

get_template_filesB

Return {filename: content} for the files that belong in a new paper project.

Excludes metadata.json and reference-only PDFs. Binary assets (images) are
base64-encoded; everything else is plain text.

Inputs: template_id (required), user_library_dir. Returns: object

humanization_promptC

Stage 3 instructions: batched parallel paragraph rewrite + correction pass.

Inputs: project_dir (required), batch_size, max_retries. Returns: result

list_directoryA

List every file under a directory, recursively, as absolute paths.

Used to discover what's actually in a research folder without already knowing the file names -- directories themselves are not included, only files.

Inputs: path (required). Returns: result

list_templatesA

Return metadata for every template in the library (seed + user-added).

Optionally filtered by paper_type ("conference", "journal", "survey").

Inputs: paper_type, user_library_dir. Returns: result

read_fileA

Read a text file at any local path on the machine this server runs on.

For a client that has no local filesystem access of its own (e.g. a browser-based chat session reaching this server through a tunnel), this is how it reads research materials, main.tex, or anything else -- a client with its own file tools (Claude Code) can use those instead, but this must work the same way for clients that can't.

Inputs: path (required). Returns: result

read_imageA

Read an image file (a figure, or a rendered PDF page) so the calling client can actually see it, not just receive a path string.

Only for genuine image files -- for a PDF page, call render_pdf_pages
first to get PNG paths, then read those here.

Inputs: path (required).

read_project_fileB

Read a plain, user-visible file previously written at the project root.

Inputs: filename (required), project_dir (required). Returns: result

read_resourceB

Read a previously written piece of pipeline handoff state.

Inputs: project_dir (required), resource_name (required). Returns: result

render_pdf_pagesA

Render every page of pdf_path to a PNG, one file per page.

Clears any previously rendered pages first so a shorter re-render
never leaves stale pages behind from an earlier iteration.

Inputs: pdf_path (required). Returns: result

screening_promptD

Stage 0 instructions: pre-flight screening of a research folder.

Inputs: project_dir (required), research_folder (required). Returns: result

structure_drafting_promptD

Stage 2 instructions: draft the paper, then compile/inspect/fix until clean.

Inputs: project_dir (required), template_id (required), research_folder (required), max_retries. Returns: result

venue_resolution_promptD

Stage 1 instructions: paper type/venue capture and template resolution.

Inputs: project_dir (required), domain_summary (required). Returns: result

write_fileB

Write (creating or overwriting) a text file at any local path.

Inputs: path (required), content (required). Returns: result

write_project_fileA

Create or fully overwrite a plain, user-visible file at the project root.

Unlike write_resource (hidden internal handoff state under
.pipeline_state/), this is for deliverables the scholar is meant to read
directly, e.g. venue.md or approach.md.

Inputs: content (required), filename (required), project_dir (required). Returns: result

write_resourceC

Write a piece of pipeline handoff state (e.g. a screening report) for a project.

Inputs: content (required), project_dir (required), resource_name (required). Returns: result

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    C2.7/5.0

    Scored across 21 tools

    Disambiguation3/5

    Several file-related tools overlap in purpose: read_file, read_project_file, and read_resource, as well as write_file, write_project_file, and write_resource, have similar semantics and only descriptive context separates them. The prompt and template tools are more clearly distinct.

    Naming Consistency3/5

    Most tools use snake_case and many follow verb_noun (e.g., list_directory, read_file, write_resource), but several pipeline prompt tools are bare noun phrases (screening_prompt, venue_resolution_prompt, structure_drafting_prompt, humanization_prompt, final_qa_prompt), breaking the consistent action-oriented pattern.

    Tool Count3/5

    With 21 tools, the surface falls in the 16-25 'heavy' range. The count is inflated by exposing each pipeline stage prompt as a separate tool rather than one parameterized prompt tool, though the tools are still coherent for the server's purpose.

    Completeness3/5

    The toolset covers file, template, rendering, and pipeline handoff operations well, but notable lifecycle operations are missing: no delete/removal for project files, templates, or resources, and template management is limited to listing, getting, and adding.

    Maintenance

    ActivityInactive
    ResponsivenessOutdated data

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      C
      quality
      B
      maintenance
      Enables academic literature management through PDF import, hybrid search, knowledge graph construction, and automated literature review generation. Combines full-text search with semantic vector search for comprehensive paper analysis.
      55
      -
    • A
      license
      A
      quality
      C
      maintenance
      Provides local academic literature search and writing support by querying OpenAlex, arXiv, and Crossref, downloading open-access PDFs, extracting IMRaD sections, and appending BibTeX references.
      4
      2
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables end-to-end research paper workflows from ideation to delivery, including paper reproduction, guided writing, export to LaTeX/PDF/DOCX, and scientific data queries across 41 public database connectors, all operable locally without API keys.
      8
      Apache 2.0