Code of Paper

Code of Paper CLI

Maintenance mode. This CLI and MCP server are not actively developed. They keep working as long as api.codeofpaper.com is up; issues may go unanswered.

Discover GitHub implementations of research papers — from the terminal.

Search 181k+ arXiv papers and find their code on GitHub. Works for humans and AI agents alike.

Install

pip install codeofpaper

Or with pipx for isolated installs:

pipx install codeofpaper

Requires Python 3.10+. No API key needed — works anonymously at 60 requests/minute.

Quick Start

# Search for papers

codeofpaper search "vision transformers"

Get paper details (arXiv IDs or full URLs both work)

codeofpaper paper 2010.11929 codeofpaper paper https://arxiv.org/abs/2010.11929

Find code implementations

codeofpaper code 1706.03762

Browse trending papers with code

codeofpaper trending --has-code

JSON output for scripts and agents

codeofpaper -o json trending | jq '.trending[] | {title, stars: .max_stars}'

One-shot reproducibility context for an agent: paper + confident repos

(with tier + framework + license) + fork-graph in a single JSON payload.

codeofpaper -o json paper 2010.11929 | jq '{title, repos: [.repos[] | {full_name, tier}], forks: [.fork_graph[] | .full_name]}'

Literature review in one command

codeofpaper research "reinforcement learning" --depth deep

Output Formats

Use -o / --output before the subcommand:

codeofpaper -o json paper 2010.11929

codeofpaper -o quiet search "attention" | head -5 codeofpaper -o csv trending --has-code > papers.csv

The shorthand -q is equivalent to -o quiet:

codeofpaper -q search "transformers" | head -3

You can set a default format via config so you don't need -o every time:

codeofpaper auth setup    # then edit ~/.config/codeofpaper/config.json

Commands

Discovery

Browsing

Bulk Operations

Configuration

Command Details

search

codeofpaper search "reinforcement learning" --sort has_code --has-code

codeofpaper search "GAN" --limit 20 --offset 10

Options: --limit, --offset, --sort (relevant, recent, has_code), --has-code, --category.

research

Multi-step orchestrated overview of a research topic:

codeofpaper research "vision transformers" --depth deep

batch

Process multiple queries or IDs from a file or stdin. Always outputs JSONL regardless of -o:

# From a file

codeofpaper batch paper ids.txt

From stdin

echo -e "2010.11929\n1706.03762" | codeofpaper batch paper

Pipe from another command

codeofpaper -q similar 2010.11929 | codeofpaper batch code

Supported commands: paper, search, code, similar, suggest.

Each line produces one JSON object:

{"input": "2010.11929", "status": "ok", "data": {...}}

{"input": "9999.99999", "status": "error", "error": "Not found"}

Options: --delay (seconds between calls, default 0.5).

export

Paginated bulk export from trending, conference, or search:

codeofpaper export trending --category cs.CV --has-code -o csv > cv.csv

codeofpaper export conference neurips_2024 --has-code -o bibtex > neurips.bib codeofpaper export search "transformers" --max 500 -o jsonl > data.jsonl

Auto-paginates through results (100 per page, 0.5s delay). Options: --max (default 200), --has-code, --category, --days.

Agent Integration

The CLI is designed for AI agent consumption. Key features:

Stable Exit Codes

Agents can branch on exit codes without parsing error messages:

Machine-Readable Output

# JSON for structured parsing

codeofpaper -o json paper 2010.11929 | jq '.title'

Quiet mode for ID lists

codeofpaper -q search "attention" | head -5

JSONL for streaming

codeofpaper -o jsonl trending --has-code | while read -r line; do echo "$line" | jq -r '.arxiv_id' done

HTTP Cache

Responses are cached on disk for 30 minutes (~/.cache/codeofpaper/http/), so repeated calls are instant and free.

MCP Server (optional)

Code of Paper ships an optional Model Context Protocol server that exposes paper / code lookup as tools to any MCP-compatible agent — Claude Desktop, Cursor, Continue, Cline, Zed, etc.

Install with the mcp extra:

pip install 'codeofpaper[mcp]'

or, with uv:

uv tool install 'codeofpaper[mcp]'

This adds a codeofpaper-mcp entry point that speaks MCP over stdio.

Wire into Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS), %APPDATA%\Claude\claude_desktop_config.json (Windows), or ~/.config/Claude/claude_desktop_config.json (Linux):

{

"mcpServers": { "codeofpaper": { "command": "uvx", "args": ["--from", "codeofpaper[mcp]", "codeofpaper-mcp"] } } }

This pulls + runs the server on demand — no global install needed. If you've already pip install'd the package, you can replace the command with just "codeofpaper-mcp" and drop args.

Wire into Cursor / Continue / Cline

Most MCP clients accept the same shape:

{

"mcpServers": { "codeofpaper": { "command": "codeofpaper-mcp", "env": { "CODEOFPAPER_API_URL": "https://api.codeofpaper.com" } } } }

Tools exposed

All tools are read-only, return plain JSON, and surface API errors as {"error": "...", "status": N} rather than throwing — agents get structured output either way.

Common Workflows

Core Discovery

# Find papers with code about a topic

codeofpaper search "reinforcement learning" --sort has_code --has-code

Get the best repo for a specific paper

codeofpaper code 1706.03762

Accepts arXiv URLs — no need to extract the ID

codeofpaper paper https://arxiv.org/abs/2010.11929

Reverse lookup: what paper does this repo implement?

codeofpaper repo google-research/vision_transformer

Open a paper in the browser, or jump straight to its code

codeofpaper open 2010.11929 codeofpaper open 2010.11929 --code

Browsing & Monitoring

# What's trending in computer vision?

codeofpaper trending --category cs.CV --sort hot

Conference papers that just got new code

codeofpaper code-drops --days 7

NeurIPS 2024 oral papers with code

codeofpaper conference neurips_2024 --track oral --has-code

Daily monitoring: what's new in my field?

codeofpaper trending --category cs.CV --days 1 -o json >> ~/research/daily_cv.jsonl

Composable Pipelines

# Find implementations of a paper's related work

codeofpaper similar 2010.11929 -o quiet | codeofpaper batch code

Batch process a reading list

cat reading_list.txt | codeofpaper batch paper > enriched.jsonl

Cross-reference with GitHub CLI

codeofpaper code 1706.03762 -o quiet | head -1 | xargs gh repo view

Export conference papers as BibTeX

codeofpaper export conference neurips_2024 --has-code -o bibtex > neurips2024_code.bib

Bulk export a category for meta-analysis

codeofpaper export trending --category cs.CV --has-code --days 365 -o csv > cv_with_code.csv

Multi-Step Scripts

# Find a paper → get top repo → clone it

REPO=$(codeofpaper search "attention is all you need" -o quiet | head -1 | \ xargs codeofpaper code -o json | jq -r '.top_repos[0].full_name') gh repo clone "$REPO"

Discover a random paper and explore its neighborhood

ID=$(codeofpaper random --quality high -o quiet) codeofpaper paper "$ID" codeofpaper similar "$ID" codeofpaper code "$ID"

Quick paper lookup in a script

ARXIV_ID="2010.11929" TITLE=$(codeofpaper -o json paper "$ARXIV_ID" | jq -r '.title') echo "Paper: $TITLE"

Configuration

Config is stored as JSON via platformdirs:

Fields:

{

"api_url": "https://api.codeofpaper.com", "api_key": null, "default_format": "table", "ca_bundle": null }

No config file needed — everything works out of the box with sensible defaults.

Priority Chain

Options are resolved in this order (first wins):

1. CLI flag (--api-url, --api-key, --ca-bundle, -o)
2. Environment variable (CODEOFPAPER_API_URL, CODEOFPAPER_API_KEY, `CODEOFP