Agent Catalog

A registry for AI agents. Track every agent you own — what it does, what tools it uses, what model it runs, where it's deployed. YAML manifests stored on disk. No database. Git-ops friendly.

Think npm for agents. One tool to list, search, diff, and invoke them all.

pip install agent-catalog
agent-catalog register ./my-agent.yaml
agent-catalog list

Why

You have agents scattered across repos, environments, and projects. Prod vs staging drift goes unnoticed. A dependency changes and nothing tells you. You grep through directories to find which agent has the capability you need.

Agent Catalog fixes that. One directory is the source of truth. Every command queries it.

Features

• Registry. Register agents from YAML manifests. List, search, get details. Filter by environment, capability, tool, surface.
• Diff. Compare manifests across environments (--left-env staging --right-env production). Catch drift before it hits prod.
• Consistency. agent-catalog doctor finds orphaned files and missing manifests.
• Security audit. Flags MCP endpoints without auth, write capabilities without confirmation gates, idempotent tools that should not be.
• MCP server. Expose the catalog as MCP tools — any MCP client can discover agents and invoke their capabilities.
• HTTP dashboard. agent-catalog serve starts a web UI on :8420.
• Python SDK. from agent_catalog.client import CatalogClient — sync or async.
• Decorator API. Define agents inline with @agent, @capability, @tool. Manifests auto-generated from Python classes.
• Runtime loader. Load an agent class from the catalog and invoke a capability from the command line or API.

Quick start

pip install agent-catalog

# Register an agent
agent-catalog register ./examples/agentic-inbox.yaml

# See everything you have
agent-catalog list

# Get details
agent-catalog get agentic-inbox

# Search
agent-catalog search --capability send_email

# Validate a manifest without registering
agent-catalog validate ./my-agent.yaml

Walkthrough: managing agents across environments

# Register your staging agent
agent-catalog register ./staging/agent.yaml --env staging

# Later, register production
agent-catalog register ./prod/agent.yaml

# See the drift
agent-catalog diff agentic-inbox --left-env staging --right-env production

# Validate your changes
agent-catalog validate ./prod/agent.yaml

# Check everything is consistent
agent-catalog doctor

Walkthrough: exposing agents via MCP

# Start the MCP server
agent-catalog serve --mcp

# Any MCP client can now call:
# - catalog_list_agents
# - catalog_get_agent
# - catalog_search
# - catalog_invoke

Walkthrough: Python SDK

from agent_catalog.client import CatalogClient

client = CatalogClient()

# List all production agents
for agent in client.agents.list(env="production"):
    print(agent.slug, agent.capability_ids())

# Async
from agent_catalog.client import AsyncCatalogClient
client = AsyncCatalogClient()
agents = await client.agents.list()

Walkthrough: decorator API

from agent_catalog import agent, capability, tool, build_manifest

@agent(name="My Agent", version="1.0.0", environment="production")
class MyAgent:
    @capability(id="greet", description="Greets the user")
    @tool(name="greet", description="Say hello")
    def greet(self, name: str) -> str:
        return f"Hello, {name}!"

manifest = build_manifest(MyAgent)

Commands

Storage

Manifests are YAML files in ~/.config/agent-catalog/agents/. index.yaml maps slugs to filenames. Set AGENT_CATALOG_DIR to point at a git repo — changes show up in git diff.

~/.config/agent-catalog/agents/
├── index.yaml            # slug → filename mapping
├── agentic-inbox.yaml    # agent manifests
├── nexusgate.yaml
└── trading-agent.yaml

Production features

• Atomic writes. Every write goes to a temp file then os.replace() — no partial writes on crash.
• Slug validation. Rejects path traversal characters (/, \, .., null).
• Index cache. Repeated index() calls within one invocation skip re-parsing.
• Config validation. Warns on type errors and unknown keys in config.yaml.
• Consistency check. doctor command finds orphaned files and missing manifests.
• XSS-safe dashboard. All agent values HTML-escaped.
• CORS headers. JSON API supports cross-origin requests.
• Graceful shutdown. SIGINT/SIGTERM signal handlers on HTTP server.
• MCP token budget. Tool responses hard-truncated to 1500 chars.

Contributing

See CONTRIBUTING.md.

License

MIT