OpenClaw MCP & Tool Integrations

What Is MCP and How OpenClaw Implements It

The Model Context Protocol standardizes how an AI client discovers and calls external tools, resources, and prompts. In OpenClaw, a saved server can use stdio or a remote HTTP transport. The CLI can inspect configuration without connecting, or probe a server to verify its live capabilities.

openclaw mcp status --verbose
openclaw mcp doctor --probe

Saving a definition does not prove that it launches or that its tools are safe. Use include and exclude filters, narrow working directories, SecretRefs, OAuth where supported, and a live probe before enabling the server for agent runs.

Connect Obsidian as a Knowledge Base for Your Agent

OpenClaw also has an official wiki bridge for Obsidian-oriented memory workflows. For a generic MCP route, the simplest option is a filesystem server scoped to one vault directory:

openclaw mcp add obsidian-vault \
 --command npx \
 --arg -y \
 --arg @modelcontextprotocol/server-filesystem \
 --arg "$HOME/Documents/MyVault" \
 --include 'read_file,list_directory,search_files'

openclaw mcp doctor obsidian-vault --probe

This example exposes files, not the full Obsidian application API. Start read-only and exclude private vault folders. If you need Obsidian-specific search, links, or commands, use the documented wiki bridge or a separately vetted server.

1Password Credential Access via MCP

OpenClaw's documented 1Password paths are SecretRefs through the op CLI and a bundled 1Password Skill. A generic MCP server is not required. SecretRef resolution keeps API keys out of openclaw.json, while the Skill supports agent workflows such as op run and op inject.

Do not give an agent broad vault access merely because MCP makes a tool easy to register. Use a dedicated service account or tightly scoped vault items, preserve 1Password audit logs, and avoid returning secrets into chat transcripts or tool output.

Composio App Integrations for OpenClaw

Composio can expose app tools through its own integration and MCP surfaces, but the number and availability of connected apps can change. Treat catalog size as a changeable product detail to verify on Composio's current official documentation, not as an OpenClaw guarantee.

For a remote MCP endpoint, save the exact HTTPS URL, prefer OAuth when supported, restrict the tool list, and run a live probe:

openclaw mcp add composio \
 --url https://mcp.example.com/mcp \
 --transport streamable-http \
 --auth oauth \
 --include 'read_*,search_*'

openclaw mcp doctor composio --probe

The URL is an example placeholder, not a Composio endpoint. Replace it with the endpoint issued for your account.

Memory Persistence With Mem0 or Cognee Plugins

Mem0 and Cognee are third-party memory products; they are not OpenClaw's default memory store. If a plugin or MCP server connects either one, verify the maintainer, transport, data retention, deletion behavior, tenancy, and write tools before use.

OpenClaw already maintains agent workspaces and session stores, and it has documented memory and wiki options. Add an external memory layer only when its retrieval or sharing model solves a defined requirement. Duplicate memory systems can create stale facts, conflicting deletion paths, and larger privacy exposure.

Tailscale for Secure Remote OpenClaw Access

Tailscale solves network reachability, not tool definition. Use it to keep a remote Gateway or MCP endpoint off the public internet, then retain Gateway authentication and MCP-level credentials. Tailscale Serve can publish the Control UI inside a tailnet; Funnel creates public exposure and is flagged as critical by OpenClaw's security audit.

For remote MCP service URLs, use TLS where possible and validate the server identity. Do not embed bearer tokens in URLs or commit literal authorization headers to configuration.

Building a Custom MCP Tool From Scratch

Build a small MCP server with a narrow schema and no ambient filesystem or shell access. Then register the local command:

openclaw mcp add local-tools \
 --command node \
 --arg ./dist/mcp-server.js \
 --cwd /srv/openclaw-tools \
 --include 'lookup_status,read_report'

openclaw mcp probe local-tools --json

The server should validate every input, return bounded output, separate read and write methods, and require confirmation for irreversible actions. doctor can check the executable and working directory; only probe establishes a live connection and lists advertised capabilities.

For browser-only tasks, compare the OpenClaw Chrome relay. For multiple isolated tool sets, continue to OpenClaw multi-agent configuration.

In Verdent

Verdent and OpenClaw are separate applications that can each participate in tool-based development workflows. Do not describe a Verdent MCP connection as an Anthropic or OpenClaw native feature unless current product documentation explicitly confirms it.

Sources: OpenClaw MCP CLI, 1Password integration, memory wiki, and OpenClaw security guide.