researchbot 0.1.1

Git-based AI research assistant with literature analysis tools

Researchbot

A Git-based AI research assistant that turns your local file system into an active research lab.

You brain dump ideas, paper references, and notes into a workspace file. Then you tell the agent what to do — survey the landscape, check novelty, deep-read specific papers — and it orchestrates the work: searching APIs, analyzing papers in parallel, and writing structured output to your project files.

Built on Claude Code with custom MCP servers, using VS Code/Cursor as the interface and Markdown as the document format.

Installation

From PyPI

pip install researchbot

From source

git clone https://github.com/juankost/researchbot.git
cd researchbot
uv sync

Configuration

Researchbot looks for API keys in this order:

1. Shell environment variables (highest priority)
2. ~/.config/researchbot/.env (user-level config)
3. <project_root>/.env (local dev fallback)

Set up your API keys:

mkdir -p ~/.config/researchbot
cat > ~/.config/researchbot/.env << 'EOF'
# Semantic Scholar (optional — increases rate limits)
SEMANTIC_SCHOLAR_API_KEY=

# Mistral OCR (required if using mistral as OCR provider)
MISTRAL_API_KEY=

# Deepseek OCR (required if using deepseek as OCR provider)
DEEPSEEK_API_KEY=

# OCR provider: "mistral" (default) or "deepseek"
RESEARCHBOT_OCR_PROVIDER=mistral
EOF

Claude Code Integration

To use researchbot as an MCP server in Claude Code globally, see claude-config for the install script that sets up skills and MCP tools.

Current Status

v1 — In Development

The project is in early development. The Semantic Scholar API client, PDF download + OCR pipeline, CLI commands, and MCP server are implemented. The agent skills are being built next.

What's implemented

• Python package structure (researchbot/) with module stubs
• Dependencies: httpx, click, mistralai, openai, mcp, python-dotenv
• Configuration with OCR provider toggle (Mistral default, Deepseek configurable), two-tier .env loading
• Semantic Scholar API client (researchbot/scholar.py) — Paper dataclass, SemanticScholarClient class with search, get_paper, citations, references, and similar paper methods. Includes resolve_paper() for flexible input: accepts IDs, URLs, or paper names.
• PDF download + OCR pipeline (researchbot/pdf.py, researchbot/ocr.py) — resolve PDF URLs via Semantic Scholar, download with caching, Mistral OCR with image extraction, per-paper cache (text.md + images/)
• CLI commands — search, paper, citations, references, similar, read with JSON output, --pretty flag, and --include-images for the read command
• MCP server (researchbot/mcp_server.py) — FastMCP server exposing 7 tools (search_papers, get_paper, get_citations, get_references, search_similar, read_paper, ocr_local_pdf) over stdio transport, auto-started by Claude Code via .mcp.json
• Skills:
  • /analyze — In-depth structured analysis of a single paper (contribution, methodology, results, limitations, future work)
  • /compare — Compare two papers on problem formulation and methodology
  • /expand — Find papers solving the same problem as a seed paper, with parallel subagent analysis
  • /gaps — Identify research gaps and open questions from a related works analysis
  • /verify_gaps — Verify which gaps are genuinely open by searching for papers that address them
  • /paper_review — Constructive conference-style review with literature verification via parallel subagents

What's next

• v1 skills complete. See docs/Vision.md for future directions.

See docs/plan.md for the full implementation plan.

Usage

CLI

# Search for papers
researchbot search "state space models" --limit 5 --pretty

# Get details for a specific paper
researchbot paper "ARXIV:1706.03762" --pretty

# Get papers that cite a paper
researchbot citations "ARXIV:1706.03762" --limit 5 --pretty

# Get papers referenced by a paper
researchbot references "ARXIV:1706.03762" --limit 5 --pretty

# Find similar papers
researchbot similar "ARXIV:1706.03762" --limit 5 --pretty

# Download + OCR a paper (markdown text output)
researchbot read "ARXIV:2106.15928"

# Download + OCR with image paths (JSON output)
researchbot read "ARXIV:2106.15928" --include-images --pretty

All commands output JSON by default. Add --pretty for indented output and --limit N to control result count.

Supported paper ID formats: ARXIV:xxx, DOI:xxx, CorpusId:xxx, Semantic Scholar hash, or URL:xxx.

The read command (and the MCP read_paper tool) also accepts paper names — they are resolved via Semantic Scholar search:

researchbot read "Attention is All You Need"

Skills (Claude Code slash commands)

Skills are invoked as slash commands inside Claude Code.

# Deep-read a paper by arXiv ID
/analyze ARXIV:1706.03762

# Deep-read by paper name
/analyze Attention is All You Need

# Deep-read by URL
/analyze https://arxiv.org/abs/2106.15928

# Compare two papers
/compare Mamba S4

# Compare papers by arXiv ID
/compare ARXIV:2312.00752 ARXIV:2111.00396

# Find related works for a seed paper
/expand Mamba

# Expand from a specific paper
/expand ARXIV:2312.00752

# Identify gaps from a related works analysis
/gaps workspace/efficient-sequence-modeling/

# Verify which gaps are genuinely open
/verify_gaps workspace/efficient-sequence-modeling/

# Review a paper (published)
/paper_review ARXIV:2312.00752

# Review a local PDF draft
/paper_review ~/drafts/my-paper.pdf

# Review with specific focus
/paper_review ARXIV:2312.00752 Focus on the theoretical claims

Documentation

File	Purpose
docs/plan.md	Implementation plan — remaining tasks and their specs
docs/next_task.md	Detailed spec for the next task to implement
docs/wip.md	Work-in-progress notes for the current task
docs/Vision.md	Long-term vision and design for the full research lab
docs/project_thoughts.md	Brain dumps on vision and project direction

Project Structure

researchbot/
  __init__.py
  config.py       # API keys, OCR provider toggle, cache paths
  scholar.py      # Semantic Scholar API client
  pdf.py          # PDF download utilities
  ocr.py          # OCR pipeline (Mistral / Deepseek)
  cli.py          # CLI entry point (Click)
  mcp_server.py   # MCP server for Claude Code (FastMCP, stdio)
.claude/
  commands/
    analyze.md      # /analyze skill
    compare.md      # /compare skill
    expand.md       # /expand skill
    gaps.md         # /gaps skill
    verify_gaps.md  # /verify_gaps skill
    paper_review.md # /paper_review skill
.mcp.json         # MCP server config (auto-starts researchbot server)
docs/
  plan.md         # Implementation plan
  next_task.md    # Next task spec
  wip.md          # Work in progress

License

MIT