Decision Hub is an open-source registry and package manager for AI agent skills, built by PyMC Labs with a focus on data science, statistics, and ML workflows. Skills are modular packages of prompts and code that any AI coding agent can use — think npm, but for agent capabilities.

Publish a skill once. Install it into Claude Code, Cursor, Codex, Gemini CLI, Windsurf, or any of 40+ supported agents with a single command. Every skill is security-scanned and can ship with automated eval cases that run on publish — so you know a skill actually works before you trust your agent with it.

Decision Hub is fully open-source (MIT) — the CLI, server, and web UI are all in this repo. You can use the public registry at hub.decision.ai, or deploy your own instance on your company's infrastructure for complete control over your skill supply chain.

• You work in data science, statistics, or ML and want AI agents that actually know your tools (PyMC, Stan, scikit-learn, pandas, etc.)
• You want proven skills, not blind trust — Decision Hub's eval pipeline tests skills against real tasks before you install them
• You maintain best practices or workflows and want to share them across your team as versioned, installable packages
• You build libraries and want to ship an AI skill alongside your package so agents know how to use it correctly
• You need a private skill registry for your organization — or want to self-host the entire platform on your own infrastructure
• You care about safety — you want skills security-scanned and graded before your agents run them

• 40+ agent support — install once, symlink into Claude Code, Cursor, Codex, Windsurf, Gemini CLI, GitHub Copilot, and many more
• Publish from anywhere — point at a local directory or a GitHub URL; every SKILL.md inside is discovered and versioned automatically
• Private skills — scope skills to your GitHub org with --private; grant cross-org access selectively
• Security gauntlet — every publish is scanned for shell injection, credential exfiltration, and dangerous patterns; skills receive a trust grade (A/B/C/F)
• Automated evals — ship eval cases with your skill; they run in an isolated sandbox, an LLM judge scores the output, and results are published as a report
• Auto-tracking — publish from a GitHub URL and a tracker automatically republishes on future commits; no CI setup needed
• Self-extending agents — Decision Hub ships as a skill itself; install it into your agent and the agent can discover and install new skills mid-conversation
• Natural language search — dhub ask "I need to do Bayesian statistics" finds relevant skills instantly

Three things set Decision Hub apart from every other skill registry:

1. Eval pipeline — Other registries accept uploads and hope for the best. Decision Hub runs your skill's eval cases in isolated sandboxes on every publish. An LLM judge scores the output. You get a public report showing whether the skill actually works. This is especially critical for data science skills where a wrong statistical method or incorrect API call can silently produce bad results.

2. Fully open-source and self-hostable — Decision Hub is the only skill registry where the CLI, server, web UI, eval infrastructure, and security scanner are all open-source (MIT). You can use the public registry, or make deploy-local to run the entire platform on your own infrastructure. No vendor lock-in, full auditability, complete control over your skill supply chain.

3. Built for data science — Created by PyMC Labs, the team behind PyMC. The registry ships with skills for Bayesian modeling, statistical analysis, and ML workflows. The search, categorization, and eval pipeline are designed for technical skills where correctness matters more than convenience.

• Not an agent framework — Decision Hub doesn't run agents. It gives agents capabilities. Use it alongside your preferred agent (Claude Code, Cursor, Codex, etc.)
• Not a prompt marketplace — there's no payment layer. Skills are published freely under open-source licenses (or kept private within your org)
• Not a closed SaaS — the entire platform is MIT-licensed. There's no "enterprise tier" hiding features behind a paywall
• Not an MCP server registry — skills are prompt+code packages that agents load directly, not server processes. They complement MCP, not replace it

# Install uv (if needed) and the CLI
curl -LsSf https://astral.sh/uv/install.sh | sh && PATH="$HOME/.local/bin:$PATH" uv tool install dhub-cli

Or if you already have uv or pipx:

uv tool install dhub-cli    # via uv
pipx install dhub-cli       # via pipx

# Authenticate via GitHub
dhub login

# Search for skills in plain English
dhub ask "I need to do Bayesian statistics with PyMC"

# Install a skill into your agent
dhub install pymc-labs/pymc-modeling --agent claude-code

# Or install into all detected agents at once
dhub install pymc-labs/pymc-modeling --agent all

# Scaffold and publish your own skill
dhub init my-skill
dhub publish ./my-skill

Each skill is a directory with a SKILL.md file. YAML front matter defines metadata; the body is the system prompt injected into the agent.

---
name: my-skill                    # 1-64 chars, lowercase alphanumeric + hyphens
description: >
  What this skill does and when
  the agent should activate it.
license: MIT

runtime:                           # optional — makes the skill executable
  language: python
  entrypoint: src/main.py
  env: [OPENAI_API_KEY]
  dependencies:
    package_manager: uv
    lockfile: uv.lock

evals:                             # optional — enables automated evaluation
  agent: claude
  judge_model: claude-sonnet-4-5-20250929
---

System prompt content goes here. This is what the agent sees
when the skill is activated.

Builds on the Agent Skills spec.

Run dhub <command> --help for full usage of any command.

dhub publish ./path/to/skills                        # from a local directory
dhub publish https://github.com/org/repo             # from a GitHub URL (auto-tracks)
dhub publish https://github.com/org/repo --ref v1.0  # specific branch/tag
dhub publish ./my-skill --minor                      # version bump: --patch (default) | --minor | --major
dhub publish ./my-skill --version 2.0.0              # explicit version
dhub publish ./my-skill --private                    # org-private visibility
dhub publish https://github.com/org/repo --no-track  # skip auto-tracking

Skills are installed as symlinks into each agent's skill directory. Use --agent NAME to target one agent or --agent all for all detected agents.

Stack: Python 3.11+ / PostgreSQL / S3 / Modal / Gemini (search) / Anthropic (eval judging)

# Clone and install
git clone https://github.com/pymc-labs/decision-hub.git
cd decision-hub
uv sync --all-packages --all-extras
make install-hooks

make help          # see all available targets
make test          # run all tests (client + server + frontend)
make lint          # ruff check + format + frontend lint
make typecheck     # mypy
make fmt           # auto-fix + format

For fully isolated local development with its own database and S3:

make deploy-local    # start Postgres + MinIO + API + frontend
# Open http://localhost:5173

make local-down      # stop (data preserved)
make local-reset     # stop and destroy all data

Requires Docker Desktop.

Copy server/.env.example to server/.env.dev and fill in your values. The project has three environments controlled by DHUB_ENV (dev | prod | local). See CLAUDE.md for detailed development guidelines.

If you discover a security vulnerability, please report it responsibly via SECURITY.md. Do not open a public GitHub issue.

MIT — see LICENSE.