Academic Agent Toolkit
JhonHander/academic-agent-toolkit
Configure ferramentas de pesquisa acadêmica em todos seus agentes de IA
Ver no GitHubMCP server que instala habilidades de pesquisa acadêmica e busca de papers em múltiplos agentes de IA (Claude, Cursor, Copilot) com uma única linha de comando. Inclui suíte de pesquisa profunda, revisão de literatura, redação de artigos, planejamento de experimentos e acesso a 20+ bases acadêmicas.
Instalação
npx convez-ai JhonHander/academic-agent-toolkit1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Revisão sistemática de literatura para dissertações e teses
- Busca integrada em arXiv, PubMed, Semantic Scholar e Crossref
- Planejamento de protocolos de pesquisa e validação de reprodutibilidade
Veredito

Ferramenta completa e bem pensada para pesquisadores acadêmicos. A abordagem unificada de configuração cross-agent resolve um problema real, mas exige Python 3.11+ e familiaridade com CLI — adequado para usuários técnicos em contexto acadêmico.
README
A tradução para o português deste README ainda está na fila.
Academic Agent Toolkit (AAT)
One command to install academic research skills and paper search across all your AI coding agents.
Students and researchers use different AI agents (Claude Code, OpenCode, Cursor, Copilot, Codex). Each agent needs different config formats, different file locations, and different skill setups. AAT handles all of that so you can focus on your research.
What you get
| Component | What it does |
|---|---|
| Academic Research Suite | Deep research, lit reviews, systematic reviews, paper drafting, peer review, experiment planning |
| Experiment Agent | Study protocol design, statistical interpretation, reproducibility validation |
| Paper Search MCP | Search 20+ academic sources (arXiv, PubMed, Semantic Scholar, Crossref...) and download PDFs |
All configured across every supported agent in one command. No manual JSON/TOML editing.
Installation
Prerequisites: Python 3.11+ and uv.
If you are not a Python developer, use the uvx option. It runs the latest published version from PyPI without making you manage a virtual environment.
Pick one:
# Recommended: run directly from PyPI, no permanent install
uvx academic-agent-toolkit doctor
# Persistent install: keeps the `aat` command available in your terminal
uv tool install academic-agent-toolkit
aat doctor
# Standard pip install
pip install academic-agent-toolkit
aat doctor
After install, run:
aat install
aat setup-keys
aat verify
Quick Start
aat doctor # See what agents are detected
aat install # Install everything (skills + MCP)
aat setup-keys # Configure API keys interactively
aat verify # Confirm everything works
How it works
New users — aat install downloads Academic Research Suite and Experiment Agent automatically, creates a private .env file for your API keys, installs one canonical skill in ~/.agents/skills/academic-research-suite, and configures Paper Search MCP in every detected agent.
Existing users — AAT adopts your existing ARS installation and Paper Search MCP registrations. It skips what you already have and only manages what's missing. Use --replace-skills or --replace-mcp if you want AAT to take over an existing setup.
The .agents/ layout is the source of truth. Agents with native skill directories get symlinks to the canonical .agents skill instead of duplicated copies.
Commands
| Command | What it does |
|---|---|
aat doctor | Show environment readiness and agent detection |
aat install | Install skill adapters and MCP configs with a guided plan |
aat setup-keys | Configure Paper Search MCP API keys interactively |
aat verify | Confirm everything is in place |
aat self-check | Validate runtime prerequisites (Python, uv, ARS source, env file) |
aat repair | Re-apply the last saved installation |
aat update | Check PyPI, upgrade AAT, then re-apply the saved installation |
aat uninstall | Remove AAT-managed files safely (does not touch your own configs) |
API Keys
Paper Search MCP works without most keys, but some sources need credentials. AAT helps you set them up interactively with aat setup-keys.
| Variable | Required? | Recommended? | Where to get it |
|---|---|---|---|
PAPER_SEARCH_MCP_UNPAYWALL_EMAIL | yes | — | unpaywall.org |
PAPER_SEARCH_MCP_CORE_API_KEY | no | yes | core.ac.uk |
PAPER_SEARCH_MCP_SEMANTIC_SCHOLAR_API_KEY | no | yes | semanticscholar.org |
PAPER_SEARCH_MCP_GOOGLE_SCHOLAR_PROXY_URL | no | no | Your proxy provider |
PAPER_SEARCH_MCP_DOAJ_API_KEY | no | no | doaj.org |
PAPER_SEARCH_MCP_ZENODO_ACCESS_TOKEN | no | no | zenodo.org |
PAPER_SEARCH_MCP_IEEE_API_KEY | no | no | developer.ieee.org |
PAPER_SEARCH_MCP_ACM_API_KEY | no | no | acm.org |
Keys are stored in a single private file (~/.config/paper-search-mcp/.env by default). Paper Search MCP reads them automatically via PAPER_SEARCH_MCP_ENV_FILE. No keys are duplicated across agent configs.
Supported Agents
Skill adapters (agent-optimized routers for Academic Research Suite):
- Zed (
~/.agents/skills/native global skill) - Claude Code
- OpenCode
- Cursor
- GitHub Copilot (
~/.agents/skills/and~/.copilot/skills/global skills) - Codex
GitHub Copilot loads skills from ~/.agents/skills/ and ~/.copilot/skills/ globally, and from .github/skills/, .claude/skills/, and .agents/skills/ per project. AAT already installs the canonical skill to ~/.agents/skills/ and symlinks to ~/.copilot/skills/, so Copilot picks it up automatically. VS Code does not expose a global skills directory — it is configured for MCP only.
MCP configuration (Paper Search MCP registration):
- Claude Code
- OpenCode
- Cursor
- Codex
- VS Code (user and global)
- GitHub Copilot
- Zed
See docs/mcp-agent-matrix.md for the exact config format used per agent.
Install flags
| Flag | Purpose |
|---|---|
--no-bootstrap | Skip automatic ARS download; requires an existing source |
--ars-source PATH | Use a specific ARS source tree |
--env-file PATH | Use a custom .env file path |
--replace-skills | Back up and replace existing skill directories |
--replace-mcp | Back up and replace existing MCP entries |
--dry-run --yes | Preview the full plan without writing files |
Uninstall
aat uninstall
Removes only files AAT created. Your existing hand-written configs and skill directories are never touched. Optional flags:
--remove-envalso removes the Paper Search MCP env file--remove-managed-arsalso removes AAT's downloaded ARS source
Where files live
| What | Location |
|---|---|
| AAT config | ~/.config/academic-agent-toolkit/config.json |
| Canonical global skill | ~/.agents/skills/academic-research-suite/ |
| Zed global agent instructions block | ~/.agents/AGENTS.md |
| Managed ARS source | ~/.local/share/academic-agent-toolkit/ars/ |
| Paper Search MCP env | ~/.config/paper-search-mcp/.env (default) |
| Skill symlinks | ~/.claude/skills/, ~/.config/opencode/skills/, ~/.cursor/skills/, ~/.codex/skills/ |
Upstream projects
AAT is an integration layer. It bundles and configures these upstream projects with their permission:
- academic-research-skills by Imbad0202 — v3.10.0
- experiment-agent by Imbad0202 — v1.1.0
- paper-search-mcp by openags
Developer testing
Use two separate environments when developing AAT:
| Environment | Purpose | Command source |
|---|---|---|
| Local / sandbox | Test your current code before publishing | This checkout on your machine |
| Production | Test what real users get from PyPI | Published academic-agent-toolkit package |
Local sandbox
From the repository root:
# Run the CLI from your local source tree
uv run aat doctor
uv run aat install --dry-run --yes
uv run aat verify
Use this before publishing. It tests the code you are editing locally, even if it is not on PyPI yet.
Production check from PyPI
Use this after publishing a release:
# Run the latest published version from PyPI
uvx academic-agent-toolkit doctor
# Or install it persistently like a real user would
uv tool install --force academic-agent-toolkit
aat doctor
aat install --dry-run --yes
aat verify
Use this to confirm that PyPI users receive the expected version and behavior.
License
MIT — see LICENSE.