π ClawPhD
March 25, 2026 Β· View on GitHub
An OpenClaw Agent for research that can turn academic papers into publication-ready diagrams, posters, videos, and more. This project is based on the nano version of OpenClaw: Nanobot.
Features
- Diagram Generation β Create publication-quality academic illustrations and statistical plots from paper sections
- Figure Reference Extraction β Search influential papers and extract real figures as editable SVG + PPTX
- PDF β Markdown + Editable Figures β Convert any paper PDF to structured Markdown; export all labelled figures as PNG + SVG + drawio
- AI Paper Review β Peer-review any PDF against NeurIPS / ICLR / ICML / EuroSys / CVPR rubrics; produces narrative feedback + dimensional scores + Accept/Reject
- Paper Discovery β Proactively search and summarize trending AI papers on a schedule
- Video Explainers β Generate walkthrough videos from paper content
- Paper Websites β Turn papers into interactive web pages
- Poster Generation β Produce conference-ready posters from papers
- Code Synthesis β Extract and generate reproducible code from paper methodologies
Examples
Diagram Generation
examples/diagram_generation_command.sh
Generated results: The following images demonstrate the Agent's iterative refinement process for generating a HumanLLM framework diagram:
First Generation (Initial Output):

After 3 iterations:

These examples showcase how the Agent progressively improves diagram quality through human-in-the-loop feedback, resulting in more polished and publication-ready outputs.
Paper Website Generation
The Agent can turn academic papers into interactive web pages:
examples/page_generation_command.sh

Figure Reference Extraction
The Agent searches influential papers and extracts all labelled figures into an editable reference pack (PNG + SVG + PPTX):
examples/figure_ref_command.sh

PDF to Markdown + Editable Figures
Convert a local paper PDF into structured Markdown and export figure assets (PNG + SVG + drawio, with editable rebuild fallback):
examples/pdf2md_command.sh
Typical output folder:
~/.clawphd/workspace/outputs/pdf2md/<pdf_stem>/
AI Paper Review
Review any paper PDF against real conference rubrics β produces a structured narrative review (Synopsis, Strengths, Weaknesses, Suggestions, References) plus venue-specific dimensional scores and an Accept/Reject recommendation:
examples/ai_review_command.sh
Supported venues: NeurIPS Β· ICLR Β· ICML Β· EuroSys Β· OSDI Β· SOSP Β· CVPR Β· ICCV Β· General
Accuracy check β ICLR 2024 papers:
| Paper | Real outcome | Originality | Significance | Contribution | Overall | Decision |
|---|---|---|---|---|---|---|
| Mamba (arXiv:2312.00752) | Accepted (Spotlight) | 4 / 4 | 4 / 4 | 4 / 4 | 8 / 10 | Accept |
| SELF-RAG (arXiv:2310.11511) | Rejected from ICLR ΒΉ | 3 / 4 | 3 / 4 | 3 / 4 | 7 / 10 | Accept |
ΒΉ Note on SELF-RAG: This is a high-quality paper β it was accepted at EMNLP 2023 and has thousands of citations. Its ICLR 2024 rejection was due to a venue policy (already published elsewhere), not a reflection of paper quality. Our system correctly evaluates it as solid, well-executed research (7/10), just not at the same level of architectural breakthrough as Mamba (8/10, with three dimensions at 4/4).
The accepted paper (Mamba) scores higher on every breakthrough dimension. All three dimensions where Mamba achieves 4/4 β Originality, Significance, Contribution β are the strongest predictors of lasting architectural influence. SELF-RAG scores a uniform 3/4, reflecting solid incremental work. Full review texts: Mamba Β· SELF-RAG.
Output folder:
~/.clawphd/workspace/outputs/paper_review/<pdf_stem>/
βββ review.md # 6-section narrative + score table
βββ meta.json # venue, mode, scores, elapsed time
Quick Start
1. Install
# From source
uv pip install -e .
# Or from PyPI
pip install clawphd-ai
2. Initialize
clawphd onboard
This creates ~/.clawphd/config.json and a default workspace at ~/.clawphd/workspace/.
3. Configure API Key
Edit ~/.clawphd/config.json and add at least one LLM provider key:
{
"providers": {
// Pick one (or more):
"openrouter": { "apiKey": "sk-or-..." },
"anthropic": { "apiKey": "sk-ant-..." },
"openai": { "apiKey": "sk-..." },
"gemini": { "apiKey": "AI..." },
"deepseek": { "apiKey": "sk-..." }
},
"agents": {
"defaults": {
"model": "anthropic/claude-opus-4-5"
}
}
}
For PaperBanana diagram generation, also set the Replicate token:
export REPLICATE_API_TOKEN="r8_..."
4. Chat
# Single message
clawphd agent -m "Hello!"
# Interactive REPL
clawphd agent
CLI Reference
| Command | Description |
|---|---|
clawphd onboard | Initialize config and workspace |
clawphd agent [-m MSG] | Chat with the agent (interactive if no -m) |
clawphd gateway [-p PORT] | Start the multi-channel gateway |
clawphd status | Show config, API keys, and workspace status |
clawphd channels status | Show channel connection status |
clawphd channels login | Link WhatsApp via QR code |
clawphd cron list | List scheduled jobs |
clawphd cron add | Add a scheduled job (--every, --cron, or --at) |
clawphd cron remove <ID> | Remove a scheduled job |
clawphd cron enable <ID> | Enable / --disable a job |
clawphd cron run <ID> | Manually trigger a job |