Smithery

The marketplace for AI agents. Connect to 100K+ tools and skills instantly.

Use smithery --help and <command> --help for flags, arguments, and full examples. This skill focuses on concepts and canonical workflows.

Quick Start

# 1. Install
npm install -g @smithery/cli

# 2. Authenticate (requires human to confirm in browser)
smithery auth login

# 3. Search for MCP servers
smithery mcp search "github"

# 4. Connect to a server (URL or qualified name both work)
smithery mcp add "https://github.run.tools" --id github

# 5. Browse tools (tree view — drill into groups by passing the prefix)
smithery tool list github
smithery tool list github issues.

# 6. Inspect tool input/output JSON schema
smithery tool get github issues.create

# 7. Call a tool
smithery tool call github issues.create '{"repo": "owner/repo", "title": "Bug"}'

Core Concepts

Namespaces

A namespace is the workspace boundary for Smithery resources. Servers, connections, and skills all live in a namespace. Use one namespace per app/environment (for example, my-app-dev, my-app-prod), then set it as your active context. Canonical flow:

smithery namespace list
smithery namespace create my-app-prod
smithery namespace use my-app-prod

For namespace-specific flags and overrides, run smithery namespace --help and smithery mcp --help.

Connect (MCP Connections)

A connection is a managed, long-lived MCP session. Smithery Connect handles OAuth flow, credential storage, token refresh, and session lifecycle. Connection status model:

connected: ready to list/call tools
auth_required: human must open authorization URL
error: inspect details and retry/fix config

Canonical flow (single user-scoped connection):

smithery mcp add https://github.run.tools \
  --id user-123-github \
  --metadata '{"userId":"user-123"}'

smithery mcp list --metadata '{"userId":"user-123"}'
smithery tool list user-123-github

If CLI shows auth_required, tell your human to open the URL and then retry.

Token Scoping

Service tokens are restricted credentials for browser/mobile/agent usage. Never pass a full API key to untrusted code. Policy mental model:

A token policy is one or more constraints
In the CLI, pass one JSON object per --policy flag
Fields inside one constraint are AND-ed (more fields = narrower)
Lists and multiple constraints are OR-ed (more entries = wider)

Canonical user-scoped token:

smithery auth token --policy '{
  "namespaces": "my-app",
  "resources": "connections",
  "operations": ["read", "execute"],
  "metadata": { "userId": "user-123" },
  "ttl": "1h"
}'

Request-level Tool Restrictions (`rpcReqMatch`, experimental)

Use rpcReqMatch to restrict specific MCP JSON-RPC requests (regex by request path). Important: connection IDs are not in the JSON-RPC body, so combine:

metadata for connection-level restriction
rpcReqMatch for method/tool restriction

Canonical combined restriction:

smithery auth token --policy '{
  "resources": "connections",
  "operations": "execute",
  "metadata": { "connectionId": "my-github" },
  "rpcReqMatch": {
    "method": "tools/call",
    "params.name": "^issues\\."
  },
  "ttl": "30m"
}'

Piped Output

When output is piped, Smithery commands emit JSONL (one JSON object per line):

smithery tool list github --flat --limit 1000 | grep label

smithery-ai-cli