Research Copilot

An AI-powered desktop research assistant for scientists and academics. Literature search, data analysis, academic writing, cross-project paper memory, and project management — powered by your ChatGPT Pro / Claude Max subscription (or an API key), all in one desktop app.

Built on pi-mono (agent runtime) + Electron + React.

Signing in (READ THIS FIRST)

Research Copilot supports three auth methods and automatically prefers the cheapest working one. When multiple are configured, priority is:

ChatGPT subscription → Claude subscription → OpenAI API key → Anthropic API key

First-launch model selection follows this order; you can override it any time from the model selector.

Option 1 — Sign in with a subscription (recommended)

The fastest and most cost-predictable path. No API key needed, no metered billing surprises.

• ChatGPT Pro / Plus — click the model selector, pick a GPT-5.4 (sub) entry, sign in via OAuth. Uses the official ChatGPT subscription endpoint.
• Claude Pro / Max — click the model selector, pick a Claude … (sub) entry, sign in via OAuth. Uses the official Anthropic subscription endpoint. (Previously gated behind ENABLE_CLAUDE_SUB=1; enabled by default since 0235a3f.)

Credentials are stored in the OS keychain via pi-ai's OAuth helper and refreshed automatically.

Option 2 — Bring an API key

Open the unified settings panel (Cmd+.) and paste a key, or set it in your shell profile:

export OPENAI_API_KEY="sk-..."           # GPT-5.4, GPT-4o, o-series
export ANTHROPIC_API_KEY="sk-ant-..."    # Claude Opus / Sonnet / Haiku

Keys entered in the UI are saved to ~/.research-copilot/config.json.

Optional supporting keys

| Key | Enhances | Without it | |-----|----------|------------| | BRAVE_API_KEY | web_search tool — general web search via Brave | Falls back gracefully to arXiv-only academic search | | OPENROUTER_API_KEY | scientific-schematics skill — AI-generated diagrams | The schematics skill fails when invoked; all other skills still work |

Semantic Scholar, arXiv, OpenAlex, DBLP: used for literature search and do not require API keys. They work out of the box.

How is Research Copilot different from Claude Cowork?

Claude Cowork is Anthropic's general-purpose autonomous agent for knowledge workers — it handles file organization, document drafting, and data extraction across everyday desktop tasks.

Research Copilot is a vertical tool built specifically for academic research. The two differ in depth, not surface:

| | Claude Cowork | Research Copilot | |---|---|---| | Scope | Horizontal — any knowledge work | Vertical — academic research lifecycle | | Literature | No academic search | Multi-source search (Semantic Scholar, arXiv, OpenAlex, DBLP) with relevance scoring, coverage tracking, and citation tracing | | Paper management | Processes files you already have | Structured artifact system with DOI, bibtex, citeKey, citation counts, and relevance metadata | | Academic writing | Generic document drafting | Venue-specific templates (NeurIPS, ICML, journals), IMRAD structure, LaTeX, citation verification (never hallucinated) | | Grant writing | None | Agency-specific guidance (NSF, NIH, DOE, DARPA, NSTC) with compliance checklists | | Data analysis | Extracts data from documents | LLM-generated Python scripts with statistical modeling, matplotlib/seaborn visualization, and output manifests | | Domain skills | General capabilities | 14 pluggable research skills (scientific writing, visualization, scholar evaluation, paper revision, slides, etc.) — extensible via Markdown | | Cross-project memory | Per-conversation only | Background Paper Wiki agent that indexes every paper you touch into a local, concept-organized knowledge base shared across all your projects | | Knowledge persistence | Not specified | Artifact store, session summaries, cross-session memory, @-mention references | | Auth | Claude subscription only | ChatGPT Pro / Claude Max via OAuth or OpenAI / Anthropic API keys — priority-ordered so subscriptions are preferred automatically | | Openness | Closed-source commercial product | Open source (MIT) — fully customizable |

In short: Claude Cowork is like a smart office assistant. Research Copilot is like a lab partner who knows how to search literature, run stats, write papers, and apply for grants.

Features

AI Chat with Coding & Writing Tools

Converse with an AI research assistant that can read, write, and edit files in your workspace. It generates LaTeX manuscripts, creates publication-quality figures, runs Python analysis scripts, and manages your project files — all through natural language.

Multi-Source Literature Search

Search across Semantic Scholar, arXiv, OpenAlex, and DBLP simultaneously. Papers are scored for relevance, deduplicated, and organized in a searchable table. Quick actions let you do deep searches, fill coverage gaps, or trace citation chains.

Cross-Project Paper Wiki

A background agent that turns every paper you've ever opened into a local, concept-organized knowledge base shared across all your projects. Each paper gets a summarized wiki page; recurring concepts get their own pages with back-references to the papers that mention them. The wiki is searchable from any project via wiki_search / wiki_get / wiki_coverage tools, so the AI can recall and cite work from earlier projects without you re-feeding it context.

The wiki runs offline and is disabled by default — it consumes LLM tokens (roughly 8K–25K input / 2K–4K output per paper), so you opt in from the Settings panel and pick a model you're comfortable paying for. Subscription-backed models are recommended; an "Auto" option follows the system-wide priority (sub before API key). Identity drift across DOI/arXiv/title lookups is reconciled automatically so papers don't get reprocessed.

Extensible Skills System

Skills are lazy-loaded knowledge modules that give the AI domain expertise. The app ships with 15 builtin skills covering academic writing (paper-writing, paper-revision, research-grants, rewrite-humanize, scientific-writing, scholar-evaluation), visualization (matplotlib, seaborn, scientific-schematics, scientific-visualization), presentations (academic-marp-slides, teaching-marp-slides), research ideation (brainstorming, creative-thinking), and general coding. You can also add your own project-specific skills as plain Markdown files.

File Attachments in Chat

Attach files directly in the chat input via the paperclip button, drag & drop, or paste. Supported formats:

| Format | How it's processed | |--------|--------------------| | Images (PNG, JPEG, GIF, WebP) | Sent as vision content — the LLM sees the image visually | | Text files (CSV, MD, TXT, JSON, XML, HTML) | Read directly and injected as text into the message | | Documents (PDF, DOCX) | Converted to text via markitdown CLI (with pypdf fallback for PDF), then injected into the message |

Note: Document conversion requires markitdown (pip install markitdown[all]) or pypdf (pip install pypdf) for PDF/DOCX files. Text-based formats work out of the box with no extra dependencies.

Future plan: The underlying Anthropic API supports native PDF document blocks (preserving layout, tables, and embedded images). Once the pi-mono agent runtime adds DocumentContent support, PDF attachments will be upgraded to use native API handling instead of text extraction.

More

• Document conversion — PDF / DOCX / PPTX / XLSX → Markdown (via agent tools)
• Python data analysis — LLM-generated analysis with matplotlib/seaborn visualization
• Artifact management — notes, papers, data, web content with CRUD tools
• @-mention system — reference entities inline in chat
• Session continuity — automatic context compaction and session summaries
• Integrated terminal — run commands without leaving the app
• LLM providers — OpenAI and Anthropic, via ChatGPT Pro / Claude Max subscription OAuth or API keys, with automatic priority selection
• Unified settings panel — Cmd+. opens a single pane for models, API keys, research presets, data-analysis timeouts, and the Paper Wiki agent

Prerequisites

• Node.js >= 18
• npm >= 9
• Python 3 (optional, for data analysis and figure generation)
• macOS recommended (Linux/Windows: use the git clone method below, untested)

Getting Started

Option A: Install via npm (recommended)

npm install -g research-copilot
research-copilot

Option B: Clone from source

git clone https://github.com/daidong/PiPilot.git
cd PiPilot
npm install
npm run dev

Authentication

On first launch, open the model selector (top of the chat pane) and either sign in with ChatGPT Pro / Claude Max via OAuth, or paste an OPENAI_API_KEY / ANTHROPIC_API_KEY into the unified settings panel (Cmd+.). Everything else is optional.

See Signing in above for the full breakdown and optional supporting keys.

Build for Production

npm run build

Project Structure

app/                  # Electron desktop application
├── src/main/         # Main process (IPC handlers, app lifecycle)
├── src/preload/      # Context bridge (renderer ↔ main)
└── src/renderer/     # React UI (components, Zustand stores)

lib/                  # Research agent logic (framework-independent)
├── agents/           # Coordinator agent + prompt registry
├── commands/         # Artifact CRUD, search, enrichment
├── mentions/         # @-mention parsing and resolution
├── memory-v2/        # Artifact storage and session summaries
├── skills/           # Skills system (loader + builtin skills)
└── tools/            # Research tools (web, literature, data, convert)

shared-electron/      # Reusable Electron IPC utilities
shared-ui/            # Shared React components and stores

Adding Custom Skills

Create a Markdown file at <your-workspace>/.pi/skills/<name>/SKILL.md:

---
id: my-skill
name: My Skill
shortDescription: Brief description of what this skill does
---

Summary loaded at startup.

## Procedures
Detailed guidance loaded on demand when the skill is activated.

Skills are auto-discovered from three locations (later overrides earlier):

1. lib/skills/builtin/ — shipped with the app
2. ~/.research-pilot/skills/ — user-global
3. <workspace>/.pi/skills/ — project-specific

Configuration

Research Copilot stores its data in the workspace under .research-pilot/:

.research-pilot/
├── artifacts/          # Notes, papers, data, web content
│   ├── notes/
│   ├── papers/
│   ├── data/
│   └── web-content/
└── memory-v2/
    └── session-summaries/

Community & Support

• Discussions — questions, ideas, usage tips, and general Q&A
• Issues — bug reports and feature requests
• Wiki — setup walkthroughs, troubleshooting, and how-tos

License

MIT