Claude for Legal
anthropics/claude-for-legalAnthropic's suite of legal-workflow plugins: 151 skills across in-house practice areas, each driven by a cold-start interview and a practice profile.
A Claude Code plugin and MCP server for US and EPO patent work: MPEP rule search, prior art search, claim checks, drafting and diagrams.
Ambitious and candid about its limits, but it is beta, has a heavy local setup, and its MCP server code was not in the files we could read.
Claude Patent Creator combines 17 skills with a Python MCP server. The server is described as hybrid retrieval over the MPEP, US code and regulations, with FAISS and BM25 indexes built locally after setup downloads about 500 MB of MPEP PDFs. Skills cover MPEP search, prior art search, patent search through SerpApi or BigQuery, EPO and EPC search and analysis, a claims analyzer, a reviewer that checks 35 USC 112 style problems, a patent application creator (a 19 KB workflow from invention description to claims and specification), PCT applications, diagram generation through Graphviz, plus setup, index management, testing and troubleshooting helpers. The README says each check reports what it ran and what it skipped, that the tool will say when an idea is not new, and that it does not file anything or give legal advice. Search options: SerpApi (free key, 250 searches a month) or Google BigQuery patent data, which needs a Google Cloud project. HyDE query expansion with Anthropic or OpenAI keys is optional and off by default. The project calls itself beta.
17 skills averaging 6.9 KB, a few with reference files and scripts, backed by an MCP server and setup tooling described in the README, but no evals.
Counted from the repository file tree. How we score depth.
This skill asks for more access than most, or sends data to another service. That is usually a design choice rather than bad intent, but you should know before you install it.
allowed-tools: Bash with no command restriction, so shell runs without a prompt while those commands are active.commands/setup-patent-system.md:3A patent drafting and search toolkit: slash commands, agents and skills plus a local Python MCP server that downloads USPTO/EPO rule texts, builds a search index with PyTorch models and queries USPTO, EPO, SerpApi and Google BigQuery. The installer builds a venv, pip-installs PyTorch, registers the MCP server at user scope in ~/.claude.json and can call sudo to install Graphviz. Many commands pre-approve unrestricted Bash. Hooks only print text, and HyDE query expansion uses the Anthropic/OpenAI key only if you set HYDE_BACKEND=api. Patent search text goes to the search providers you configure.
pip install -e ., and can install Graphviz (brew, winget, or sudo apt-get/dnf install).install.py:361claude mcp add ... --scope user (~/.claude.json) and, on request, runs gcloud auth application-default login which writes Google ADC credentials under ~/.config/gcloud.install.py:591curl https://sdk.cloud.google.com | bash as manual advice for installing the Google Cloud SDK; the installer does not run it.install.py:477pip install google-auth-oauthlib without a version if the import fails; README/CLAUDE.md also show pip install git+https://... installs.scripts/setup_bigquery_auth.py:17The repository (177 files) on Oct 29, 2026. Not a full review: we read 15 of 106 code files; it relies on code outside the repository: Python packages installed by pip (torch, sentence-transformers, google-cloud-bigquery, requests, etc.) and the BAAI/bge-base-en-v1.5 model downloaded from Hugging Face; Google Cloud SDK (gcloud) and Graphviz system packages; Most of the 106 listed code files (the mcp_server/ modules including cli.py, server.py, bigquery_search.py, uspto_api.py, epo_api.py, hyde.py, graphviz_installer.py, plus tests) were checked with targeted pattern searches and excerpts, not read in full.
NVIDIA SkillSpector static scan: 206 pattern hits (Privilege Escalation 78, Dangerous Code Execution 51, Data Exfiltration 26, Supply Chain 20, Excessive Agency 13, Tool Misuse 10). The CRITICAL taint hit is the SerpApi URL built from the user's own SERPAPI_API_KEY in google_patents_search.py (key goes only to serpapi.com and is redacted from errors). The 'backdoor_persistence' YARA hits are docs telling the user to append API-key exports to ~/.bashrc. The curl|bash lines are printed advice; rm {MPEP_DIR}/*.pdf and rm -rf mcp_server/index_backup_* are cleanup instructions in docs/skills; env-copy hits pass the environment to child processes during setup.
A review, not a certification. Skills change after we check them — read the files before installing.
MIT — commercial use allowed. The README says output is not legal advice.
Not really. Drafting and search workflows are interactive and the README stresses human review before anything is filed.
Anthropic's suite of legal-workflow plugins: 151 skills across in-house practice areas, each driven by a cold-start interview and a practice profile.
A curated, contributor-written library of about 280 legal agent skills (contracts, GDPR, litigation, AI Act) in English, French, Italian and more.
Harvard MIMS lab's Python SDK, MCP server and 556 skills that give agents access to thousands of scientific databases and tools for biomedical research.
An open-source Claude Code plugin and MCP server for patents: it searches the MPEP and patent databases, checks drafts for common defects, helps draft claims and specifications, and generates patent diagrams.
Use /plugin marketplace add RobThePCGuy/Claude-Patent-Creator and /plugin install claude-patent-creator-standalone@claude-patent-creator, then run the setup command. A pip install from the GitHub repo followed by patent-creator setup is the alternative.
No. The README says it does not file anything and is not a lawyer; it prepares a draft and evidence for a registered patent practitioner to review.
Claude Code needs a Claude subscription. Patent search uses a free SerpApi key (250 searches a month) or BigQuery, which requires a Google Cloud project with billing enabled.