TheAppliedScientist

TheAppliedScientist searches literature, reviews papers, runs experiments, and revises manuscripts.

Contents

Hosted MCPs · Full setup · Your paper · Individual components · Docs

In your paper's folder, run the block for your agent. Use Bash on Linux/macOS or WSL/Git Bash on Windows.

claude mcp add --transport http appliedscientist-search https://search.eigenlabs.online/mcp
claude mcp add --transport http appliedscientist-review https://review.eigenlabs.online/mcp

mkdir -p .claude/skills/appliedscientist
curl -fsSL https://raw.githubusercontent.com/TheAppliedScientist/TheAppliedScientist/main/skills/appliedscientist/SKILL.md \
  -o .claude/skills/appliedscientist/SKILL.md

codex mcp add appliedscientist-search --url https://search.eigenlabs.online/mcp
codex mcp add appliedscientist-review --url https://review.eigenlabs.online/mcp

mkdir -p .agents/skills/appliedscientist
curl -fsSL https://raw.githubusercontent.com/TheAppliedScientist/TheAppliedScientist/main/skills/appliedscientist/SKILL.md \
  -o .agents/skills/appliedscientist/SKILL.md

No login is needed. Ignore Codex's suggestion to run codex mcp login.

Open or restart your agent in that folder and try one of these requests.

Review	Search
Use the AppliedScientist skill to review papers/draft.pdf.	Use the AppliedScientist skill to find papers on test-time adaptation.
Use the AppliedScientist skill to review paper-source.zip.	Use the AppliedScientist skill to find papers related to arXiv:1706.03762.

Your agent uploads the file and retrieves the review. The Reviewer handles its own literature searches.

Review limits

20 MB per file · first 12 rendered pages · several minutes plus queue time.

File formats · Usage skill · Search API docs · Review API docs

1. Install prerequisites and clone

Choose a runtime	Prerequisites
Docker	Python 3.12, Git, Docker Engine 27+, Compose 2.30+
Native, without Docker	Python 3.12, Node.js 22, Git, uv, and research tools

Install on Linux · macOS · Windows / WSL 2

Allow about 20 GB for Search/Reviewer or 60 GB for Docker GPU runs. Search downloads a 13 GB index once.

git clone https://github.com/TheAppliedScientist/TheAppliedScientist.git
cd TheAppliedScientist

2. Prepare API keys

Setup asks for these values and saves them in .env.

Component	Have ready
Search embeddings	A Gemini API key
Search full-paper answers	Uses the same Gemini key with gemini-3-flash-preview by default
AI Reviewer	Endpoint URL, API key, and model name
AI Scientist	Endpoint URL, API key, and model name

Reviewer and Scientist need Anthropic-compatible endpoints. OpenAI-only endpoints need a converter for Claude Code. Provider settings.

To change Search's answer model, edit SEARCH_QUERY_* in .env after setup, before startup. Examples.

3. Install components and start the APIs

Choose one runtime.

./tas setup search reviewer scientist --runtime docker
./tas doctor
./tas up
./tas smoke

Add --dockerize-search to setup if you also want Search in Docker.

./tas setup search reviewer scientist --runtime native
./tas doctor
./tas up
./tas smoke

Search runs on port 8081 and Reviewer on 8082. smoke tests Search retrieval and Reviewer health.

Add your paper below, then start TheAppliedScientist.

4. Place your files

Inside the cloned repository, open components/ai-scientist/user-ideas/. Use this folder layout.

user-ideas/
├── my-paper.json             ← the configuration you create in step 5
└── my-paper/
    ├── paper-source/         ← your complete LaTeX project
    │   ├── main.tex
    │   ├── references.bib
    │   └── figures/
    ├── code/                 ← your experiment code, strongly recommended
    └── reviews.md            ← paste all existing reviews here

Use your actual LaTeX filenames. Include any style files and other source dependencies.

If you have no reviews, omit reviews.md and set Human_Reviews_From_OpenReview to false. If you have no code, omit code/ and remove code_path.

5. Create my-paper.json

These paths start from the folder containing my-paper.json. Absolute paths work too.

{
  "Name": "my_paper",
  "Title": "My Paper",
  "Task": "Read the paper, public reviews, and code. Reproduce the main result, address the reviewer concerns with real experiments, revise the paper, and continue iterating with the AI Reviewer.",
  "Paper": {
    "local_path": "my-paper/paper-source",
    "code_path": "my-paper/code"
  },
  "Human_Reviews_From_OpenReview": {
    "source_file": "my-paper/reviews.md"
  },
  "What_NOT_To_Do": "Do not fabricate results, citations, or experiments.",
  "GPU_Needed": false
}

Your case	JSON change
No existing reviews	Set "Human_Reviews_From_OpenReview": false.
No experiment code	Remove code_path; describe what can be reproduced in Task.
GPU experiments	Set "GPU_Needed": true.

For papers without reviews, use this Task.

Read the paper and code. Reproduce the main result, use Search to find gaps, test improvements with experiments, revise the paper, and iterate with the AI Reviewer.

Other input formats and complete examples.

6. Start TheAppliedScientist

From the repository root, run the command for your runtime.

# Docker
./tas run components/ai-scientist/user-ideas/my-paper.json -- --timeout 21600

# Native, without Docker
./tas run --runtime native components/ai-scientist/user-ideas/my-paper.json -- --timeout 21600

For GPU experiments in Docker, complete NVIDIA setup, then add --gpus 1. Native mode uses host devices directly.

Run limit: six hours in these examples

Outputs

components/ai-scientist/jobs/my-paper__<timestamp>/

Experiment code · results · figures · paper source/PDF · review history · agent trajectory.

Install the prerequisites, then choose what to run.

Component	Install	Start
Search API	./tas setup search --runtime native	./tas up search
AI Reviewer	./tas setup reviewer --runtime docker	./tas up reviewer
AI Scientist	./tas setup scientist --runtime docker	./tas run components/ai-scientist/user-ideas/my-paper.json

Use --runtime native for Reviewer or Scientist without Docker. Setup asks only for the selected components' credentials.

Component	Required APIs
AI Reviewer	A running Search API
AI Scientist	Running Search and Review APIs

Set the service URLs / use separate machines · Host your own public APIs and MCPs

How it works

The Scientist runs experiments and revises the paper. Each independent review guides the next revision. Code, results, and drafts carry forward.

Guide	Contents
Hosted MCPs	Supported files, usage, review limits
Installation	Linux, macOS, Windows/WSL, Docker, native, GPU
API keys and models	Gemini embeddings, Search answer models, agent endpoints
Deployment	Individual components, separate machines, service commands
Paper inputs	Source, code, reviews, public URLs
Original experiment setup	Reference prompts, versions, compatibility checks

./tas status          # service health
./tas doctor          # configuration checks
./tas logs search     # follow Search logs
./tas logs reviewer   # follow Reviewer logs
./tas down            # stop services

Docker research jobs can control Docker through its socket. Native jobs execute on the host. Use a dedicated research machine; keep .env and raw trajectories private. Publishing outputs safely.

MIT license. Bundled dependencies retain their original notices.