Install + run any registry agent in 3 commands.

Quick start: the bundled-tool path

The shortest install in the registry uses bundled-Python tools. There is no MCP server to host, no API key to set, no separate config file. One curl, one pip install, and the agent runs locally:

# 1 · download the runnable Python stub
curl -sL https://agentarium.science/a/cmr-care-local/v/1.0.0.openai-agents.py \
  -o cmr-care-local.py

# 2 · install the SDK + the bundled tool
pip install openai-agents agentarium-cmr-local

# 3 · run it
python cmr-care-local.py "Find Arctic sea ice datasets for 2023"

The downloaded file is a self-contained Python module: Agent(...) with its CARE prompt, plus tools=[search_collections, get_collection_metadata, get_granules] imported directly from the bundled CMR package. No HostedMCPTool, no .mcp.json, no Bearer token. The only network egress is to NASA's public CMR API itself.

Three ways a tool can run

Every registered tool declares a transport. The registry serves the right install artifact for each:

• http — remote MCP server reached over HTTPS. The operator hosts it; you provide an API key in your env. Example: cmr_mcp_server (FastMCP wrapper hosted by NASA-IMPACT).
• stdio — a local subprocess speaking MCP over stdio. The agent runtime spawns it via { command, args }. Example: playwright_local (npx @playwright/mcp@latest).
• bundled-python — a Python module the agent imports directly. No MCP layer, no subprocess, no network roundtrip. Example: cmr_local_python (in-process wrapper around NASA's open CMR REST API).

Install paths by runtime

Each agent page has an INSTALL section with copy-able commands per client. Here's what each runtime gets and what it doesn't:

Claude Code

Subagent file + MCP config. Two artifacts, both copy-pasteable:

# 1 · drop the subagent file into your Claude Code agents dir
curl -sL https://agentarium.science/a/<slug>/v/<version>.md \
  -o ~/.claude/agents/<slug>.md

# 2 · register the MCP servers (per-agent, http or stdio tools)
claude mcp add --transport http <name> <url>          # for http tools
claude mcp add <name> --command <cmd> --args <args>   # for stdio tools

The .md file's frontmatter declares the agent's preferred model. Claude Code substitutes its own model when running; the routing and tool calls still work as advertised. Bundled-Python tools do not work in Claude Code without a stdio wrapper — the format can't import a Python module as a tool.

Cursor

Project rule + MCP fragment. Cursor reads .mdc rules from .cursor/rules/ and MCP servers from .cursor/mcp.json:

# 1 · drop the project rule
curl -sL https://agentarium.science/a/<slug>/v/<version>.cursor.mdc \
  -o .cursor/rules/<slug>.mdc

# 2 · merge the MCP-server fragment into your existing mcp.json
curl -sL https://agentarium.science/a/<slug>/v/<version>.cursor.mcp.json \
  -o .cursor/mcp.json.new

Cursor has no per-agent tool-approval gate, so author-declared user-approval guardrails in the prompt are advisory only — not enforced at runtime. Same bundled-Python caveat as Claude Code.

OpenAI Agents SDK

The richest install. Handles all three transports — http via HostedMCPTool, stdio via MCPServerStdio, bundled-Python via direct from … import:

# 1 · install the SDK
pip install openai-agents

# 2 · download the runnable Python stub
curl -sL https://agentarium.science/a/<slug>/v/<version>.openai-agents.py \
  -o <slug>.py

# 3 · install bundled tools + set env vars (printed in the file's docstring)
pip install <each-bundled-tool>
export <each-required-env-var>=...

# 4 · run
python <slug>.py "your query"

The rendered file's docstring lists every one-time setup command needed — pip installs for bundled tools, npm installs for stdio tools, env-var names for http tools. Credentials never appear in the file; it only references os.environ[...].

ChatGPT (Custom GPT)

Prompt only. Copy the agent's prompt into a Custom GPT's Instructions field:

curl -sL https://agentarium.science/a/<slug>/v/<version>/prompt.md

Custom GPTs can't attach MCP servers, so tool calls will not execute. You get the reasoning scaffold without the live data connection.

Where credentials live

Never in the registry. Every tool record has a How to get a key panel on its detail page, listing: who to email or which portal to request from, what env vars to set in your agent runtime, and eligibility (if any). The registry never sees, stores, or proxies the keys you obtain.

• http + api-key — set the *_KEY env var declared by the tool. The.openai-agents.py file references it via os.environ["FOO_KEY"].
• http + open — no credential. Public read-only APIs (e.g. cmr_mcp_server).
• stdio — credentials, if needed, are set the same way you'd set them for any local tool the subprocess inherits.
• bundled-python — credentials live in your Python process's env, read by the bundled module itself.

For other agents (programmatic access)

The registry exposes the full record of every agent and tool via JSON. Useful for agents that want to discover other agents:

# List agents (with optional ?q=, ?domain=, ?task=, ?min_tier= filters)
GET https://api.agentarium.science/api/v1/agents

# One agent — full record with prompt, required_tools, evaluations
GET https://api.agentarium.science/api/v1/agents/<slug>

# List tools
GET https://api.agentarium.science/api/v1/tools

# One tool — record + auth + access_instructions
GET https://api.agentarium.science/api/v1/tools/<id>

# OpenAPI spec
GET https://api.agentarium.science/api/openapi.json

Listings support full-text search across name, description, task, and keywords via ?q=. Results are ranked by Postgres FTS; semantic search is on the roadmap.

Submitting your own

Sign in with ORCID, get endorsed by a moderator, then submit an agent or submit a tool. Each submission goes through a conformance gate (format, topic, reachability) — not a correctness review. See Governance for details on endorsement, moderation, and the safety screen.

Citing what you use

Every agent and tool has a citable URL shown in its header. The URL resolves to a specific version and is stable — versioning never breaks an existing citation. Drop it into your paper alongside your data and code citations.