MCP Tools

All MCP tools are available to every model — model-agnostic by design. OpenAgent includes 9 default MCPs that load automatically, and the LLM can enumerate them at runtime via the built-in list_mcp_servers tool.

Default MCPs

Name	What it does	Requires
vault	Read/write Obsidian-compatible markdown notes	Node.js
filesystem	Read, write, list, search files	Node.js
editor	Find-replace, grep, glob	Node.js
web-search	Web search + page fetch, no API key	Node.js + Playwright
shell	Cross-platform shell execution	Node.js
computer-control	Screenshot, mouse, keyboard (macOS/Linux/Windows)	Node.js
chrome-devtools	Browser automation, DOM, performance	Node.js + Chrome
messaging	Send messages via Telegram/Discord/WhatsApp	Channel tokens
scheduler	Manage cron tasks from within conversations	Python

Tool names are namespaced <server>_<tool> (Agno convention) so filesystem_read_text_file, vault_write_note, etc. — no collisions between servers, and the LLM can pick tools just by name.

Disabling Defaults

mcp_defaults: false                    # disable all
mcp_disable: ["computer-control"]      # disable specific ones

Adding Your Own MCPs

User MCPs are merged on top of defaults. Same name = override.

mcp:
  # Any stdio MCP
  - name: github
    command: ["github-mcp-server", "stdio"]
    env:
      GITHUB_PERSONAL_ACCESS_TOKEN: ${GITHUB_TOKEN}

  # npx-based MCPs
  - name: sentry
    command: ["npx", "-y", "@sentry/mcp-server@latest"]
    env:
      SENTRY_ACCESS_TOKEN: ${SENTRY_TOKEN}

  # Remote MCP (SSE or HTTP)
  - name: remote-tool
    url: "http://localhost:8080/sse"

  # Remote MCP with OAuth
  - name: service
    url: "https://mcp.example.com/sse"
    oauth: true

How the Pool Works

At startup, OpenAgent builds a single MCPPool that connects every configured server once. Both model backends read from the same pool:

• Agno providers (OpenAI, Anthropic API, Z.ai GLM, any OpenAI-compatible endpoint) get the live MCPTools toolkits and register them on the Agno Agent directly — tool routing, call loops, and retries are Agno's native implementation.
• Claude CLI receives the stdio/URL spec dicts and hands them to the Claude Agent SDK's ClaudeSDKClient(mcp_servers=...) parameter, which manages its own subprocess lifecycle.

Sharing the pool means you don't pay N times for the same MCP when running SmartRouter across multiple tiers, and there's no in-process tool registry for OpenAgent to keep in sync — the providers own it.

Per-call timeout defaults to 30 seconds (npx-launched servers like @modelcontextprotocol/server-filesystem routinely take 5–15 s on their first tool call).

Claude CLI Notes

When using claude-cli as the model provider, OpenAgent passes the same pool to the Claude Agent SDK via mcp_servers. Known constraints:

• Requires Claude CLI 2.1.96+ and the claude-agent-sdk Python package
• Commands must use absolute paths (handled automatically)
• The Claude binary is occasionally unreliable when many MCPs register at once; OpenAgent sets --strict-mcp-config so failures surface immediately instead of silently dropping servers